Анализ сайта khoadoan.me
Основное Готовность: 75%
Домен
khoadoan.me
Состояние доменного имени
?
Проверяем корректность доменного имени и наличие технических проблем на уровне домена.
Домен второго уровня идеален для продвижения.
Отличный запоминающийся домен.
Ответ сервера
200 Успешный ответ
HTTP-код ответа и цепочка редиректов
?
Код 200 — страница доступна. Коды 3xx — редиректы (цепочки замедляют загрузку и размывают ссылочный вес). Коды 4xx/5xx — ошибки, поисковик не сможет проиндексировать страницу.
Сервер настроен корректно.
Цепочка редиректов:
http://khoadoan.me
302 Found
https://mail-research.com/
200 OK
Безопасность
Сайт безопасен
Использование HTTPS и SSL-сертификат
?
HTTPS — обязательный стандарт. Google и Яндекс отдают предпочтение защищённым сайтам. Отсутствие SSL или просроченный сертификат ведут к предупреждениям в браузере и снижению позиций.
Не настроен HSTS (Strict-Transport-Security) — рекомендуется включить.
На сайте работает защищенный протокол ssl и сайт открывается по https.
Ssl-сертификат действителен до 18.11.2026 21:44:36.
HTTP автоматически перенаправляется на HTTPS.
Поздравляем! Сайт не содержится в реестре РКН.
Кодировка
utf-8
Кодировка символов страницы
?
Стандарт — UTF-8. Неправильная кодировка вызывает нечитаемые символы и мешает поисковику корректно распознать текст страницы.
Указана кодировка на странице utf-8.
Язык
Атрибут lang в HTML-теге
?
Атрибут lang (<html lang="ru">) сообщает поисковикам и браузерам, на каком языке написана страница. Помогает при ранжировании в региональном поиске.
Язык страницы не указан. Рекомендуется явно указать язык документа!
Скорость загрузки
~0,88сек
Время отклика сервера (TTFB)
?
Time To First Byte — время до получения первого байта от сервера. Норма до 200 мс. Медленный отклик ухудшает пользовательский опыт и ранжирование: Яндекс и Google учитывают скорость страниц.
Скорость загрузки сайта 0,88сек оптимальна.
Объем документа
223Кб
Размер HTML-кода страницы
?
Слишком большой HTML замедляет парсинг браузером и сканирование поисковым роботом. Рекомендуется не более 200 Кб.
Объем html-документа 223Кб оптимален.
Структура html-документа корректна.
Ресурсы
Ресурсы: 24
Внешние ресурсы страницы (CSS, JS, изображения)
?
Количество и тип подключённых ресурсов влияют на скорость загрузки. Большое число запросов увеличивает время рендеринга страницы.
Кол-во файлов ресурсов 24 много для одной страницы. Приемлемо до 10. Проведите оптимизацию файлов ресурсов!
Показать полный список ресурсов
| Тип | Название | Значение |
|---|---|---|
| stylesheet | https://stackpath.bootstrapcdn.com/bootstrap/4.5.2/css/bootstrap.min.css | |
| stylesheet | https://cdnjs.cloudflare.com/ajax/libs/mdbootstrap/4.19.1/css/mdb.min.css | |
| stylesheet | https://cdn.jsdelivr.net/npm/bootstrap-icons@1.10.5/font/bootstrap-icons.css | |
| stylesheet | https://cdnjs.cloudflare.com/ajax/libs/font-awesome/5.14.0/css/all.min.css | |
| stylesheet | https://cdnjs.cloudflare.com/ajax/libs/academicons/1.9.0/css/academicons.min.css | |
| stylesheet | text/css | https://fonts.googleapis.com/css?family=Roboto:300,400,500,700|Roboto+Slab:100,300,400,500,700|Material+Icons |
| stylesheet | assets/font-awesome-4.7.0/css/font-awesome.min.css | |
| stylesheet | https://combinatronics.io/jwarby/jekyll-pygments-themes/master/github.css | |
| stylesheet | /assets/css/main.css | |
| js | https://cdnjs.cloudflare.com/ajax/libs/jquery/3.5.1/jquery.min.js | |
| js | https://www.googletagmanager.com/gtag/js?id=G-5E8HDGL2HV | |
| js | https://platform.x.com/widgets.js | |
| js | text/javascript | https://cdn.jsdelivr.net/npm/mathjax@3.2.0/es5/tex-mml-chtml.js |
| js | https://polyfill.io/v3/polyfill.min.js?features=es6 | |
| js | text/javascript | //cdn.clustrmaps.com/map_v2.js?cl=ffffff&w=200&t=n&d=6Aazaa3W3ib-sAIfPEQYBd0A-PM2ZbkZfCUcYo54GsE&co=233d4d |
| js | https://cdnjs.cloudflare.com/ajax/libs/popper.js/2.4.4/umd/popper.min.js | |
| js | https://stackpath.bootstrapcdn.com/bootstrap/4.5.2/js/bootstrap.min.js | |
| js | https://cdnjs.cloudflare.com/ajax/libs/mdbootstrap/4.19.1/js/mdb.min.js | |
| js | https://unpkg.com/masonry-layout@4/dist/masonry.pkgd.min.js | |
| js | https://unpkg.com/imagesloaded@4/imagesloaded.pkgd.min.js | |
| js | text/javascript | /assets/js/mansory.js |
| js | https://cdn.jsdelivr.net/npm/medium-zoom@1.0.6/dist/medium-zoom.min.js | |
| js | /assets/js/zoom.js | |
| js | /assets/js/common.js |
Серверные заголовки
Кол-во: 17
HTTP-заголовки ответа сервера
?
Заголовки сервера передают браузеру и поисковику служебную информацию: кеширование, безопасность (CSP, HSTS), сжатие (gzip). Правильная настройка ускоряет загрузку и повышает защищённость.
Найдены серверные заголовки 17шт. Подробнее про серверные заголовки.
Показать полный список серверных заголовков
| Ключ | Значение |
|---|---|
| Server | GitHub.com |
| Access-Control-Allow-Origin | * |
| ETag | "6a854ccf-37ed7" |
| Cache-Control | max-age=600 |
| x-proxy-cache | MISS |
| x-github-request-id | 93FE:12CAF9:220C560:2256000:6A89C278 |
| x-github-edge-region | fra |
| Accept-Ranges | bytes |
| Date | Sat, 22 Aug 2026 15:38:32 GMT |
| Via | 1.1 varnish |
| Age | 0 |
| X-Served-By | cache-fra-eddf8230118-FRA |
| X-Cache | HIT |
| x-cache-hits | 1 |
| x-timer | S1787413113.614843,VS0,VE112 |
| Vary | Accept-Encoding |
| x-fastly-request-id | 5e05370f04888e1a03a3f74aa99f8a7e6df23f98 |
CMS
Не определена
Система управления сайтом (движок)
?
CMS — это движок, на котором работает сайт (WordPress, 1C-Bitrix, Tilda и др.). Знание CMS помогает понять возможности SEO-оптимизации и подобрать подходящие инструменты. «Не определена» — вероятно, самописный сайт или нестандартная сборка.
CMS не определена. Вероятно, сайт самописный либо движок надёжно скрыт. Это не ошибка.
Веб-сервер
GitHub.com
Программное обеспечение сервера
?
Веб-сервер — это ПО, которое отдаёт страницы посетителям (nginx, Apache, IIS, LiteSpeed и др.). Определяется по серверным заголовкам ответа (Server, X-Powered-By и т.п.). «Не определён» — сервер намеренно скрывает эти заголовки, это нормальная практика безопасности.
В заголовке Server указано: GitHub.com.
Мета-теги Готовность: 17%
Title
Khoa D Doan
Заголовок страницы в браузере и поисковой выдаче
?
Title — главный SEO-заголовок страницы. Влияет на CTR в поиске и ранжирование. Оптимальная длина: 50–70 символов. Ключевые слова — ближе к началу.
Необходимо увеличить число символов в title (текущее значение мало: 18, минимум: 25, оптимально: от 40 до 45)
Дублей словоформ в title не найдено.
Description
A simple, whitespace theme for academics. Based on [*folio](https://github.com/bogoli/-folio) design.
Описание страницы в поисковой выдаче (сниппет)
?
Meta Description — текст под заголовком в выдаче. Напрямую на позиции не влияет, но влияет на CTR. Оптимальная длина: 120–160 символов.
Необходимо увеличить число символов в description (текущее значение: 102, оптимально: от 120 до 130)
Keywords
Список ключевых слов страницы (устаревший тег)
?
Meta Keywords не учитывается Яндексом и Google для ранжирования с 2009–2012 годов. Заполнение не обязательно, но не вредит. Конкурент может использовать содержимое для анализа.
Установите мета-тег keywords!
Канонический Url
/
Указывает поисковику основную версию страницы
?
Canonical (rel=canonical) предотвращает проблему дублей страниц. Должен точно совпадать с URL проверяемой страницы. Неправильный canonical может передать ссылочный вес на другую страницу.
Домен в каноническом Url не совпадает!
Robots
Ошибок нет
Директивы для поисковых роботов на уровне страницы
?
Meta Robots управляет индексацией конкретной страницы: index/noindex — индексировать ли, follow/nofollow — следовать ли по ссылкам. Noindex полностью исключает страницу из поиска.
Meta-тег robots не указан. Страница свободна для индексации.
Адаптивность
width=device-width, initial-scale=1, shrink-to-fit=no
Настройка масштабирования на мобильных устройствах
?
Тег viewport (<meta name="viewport">) сообщает браузеру, как масштабировать страницу на мобильных. Стандарт: width=device-width, initial-scale=1. Отсутствие — признак отсутствия мобильной версии.
Meta-тег viewport со значением-константой width=device-width задаёт ширину страницы в соответствии с размером экрана.
Meta-тег viewport со значением initial-scale=1.0 определяет масштаб 1:1, т.е. «не масштабировать».
Разметка OpenGraph
Не найдено
Мета-теги для красивых превью в соцсетях
?
OpenGraph (og:title, og:description, og:image) управляет тем, как страница выглядит при репосте в социальных сетях и мессенджерах. Отсутствие OG-тегов — невзрачный превью при шеринге.
Разметка OpenGraph не задана. Страница не оптимизирована под социальные сети. Мета-теги с разметкой Og помогают социальным роботам лучше структурировать Ваш сайт.
Все мета-теги
Кол-во: 2
Полный список мета-тегов страницы
?
Таблица всех meta-тегов, включая нестандартные. Позволяет найти опечатки, дубли и лишние теги.
Найдены мета-теги 2шт. Мета-теги не видимы для человека и предназначены для обмена информацией между веб-страницей и поисковыми системами, браузерами и другими веб-службами. С ними роботы 🤖 и устройства ведут себя более ожидаемо.
Показать полный список мета-тегов
| Тип | Название | Значение |
|---|---|---|
| name | viewport | width=device-width, initial-scale=1, shrink-to-fit=no |
| name | description | A simple, whitespace theme for academics. Based on [*folio](https://github.com/bogoli/-folio) design. |
Оптимизация Готовность: 40%
Структура
Ошибок нет
Семантические HTML-элементы страницы
?
Проверяет наличие основных структурных элементов: nav, header, footer, main. Корректная семантическая структура помогает поисковику понять архитектуру страницы.
Структура документа корректна (теги <html> и <body> присутствуют в единственном экземпляре).
Контент
Есть ошибки
Объём и качество текстового содержимого
?
Анализирует объём полезного текста на странице. Слишком мало — страница может считаться малополезной. Слишком много — ухудшается читаемость и восприятие.
Слова из title 2 встречаются в тексте редко. Добавьте в контент страницы слова из тега <title>!
Абзацев с текстом 55 достаточно.
Среднее число слов в абзаце 112 достаточно.
Кол-во знаков контента 72469 на странице оптимально.
Кол-во слов 10528 на странице оптимально.
Заголовки
Ошибок нет
Иерархия заголовков H1–H6
?
H1 должен быть один и содержать ключевой запрос. H2–H6 описывают подразделы. Пропуск уровней (H1 → H3) и несколько H1 — типичные ошибки, снижающие понятность страницы для поисковика.
На странице присутствуют заголовки <h2> 3. Это хорошо.
Тошнота
9,43
Насколько одно слово доминирует в тексте
?
Классическая тошнота = √(частота самого повторяющегося слова). Норма до 7–8: текст воспринимается естественно. Выше — поисковик может счесть страницу переспамленной.
Тошнота превышает норму 5. Измените текст страницы!
Академич. тошнота
583,89%
Насколько текст перенасыщен ключевыми словами
?
Академическая тошнота = (частота слова / общее количество слов) × 100%. Показывает долю конкретного слова в тексте. Норма 5–15%.
Академическая тошнота превышает норму 5-15%. Измените текст страницы!
Семантическое ядро
20
Наиболее часто встречающиеся слова на странице
?
Топ слов по частоте использования. Показывает, какие слова доминируют в тексте с точки зрения поисковика.
Контент страницы содержит осмысленный текст и слова.
Показать список слов
| Слово | Кол-во | Частота |
|---|---|---|
| backdoor | 89 | 0,85% |
| attacks | 56 | 0,53% |
| {true} | 54 | 0,51% |
| learning | 44 | 0,42% |
| nguyen | 44 | 0,42% |
| models | 37 | 0,35% |
| attack | 37 | 0,35% |
| neural | 32 | 0,30% |
| selected | 29 | 0,28% |
| conference | 29 | 0,28% |
| author | 27 | 0,26% |
| bibtex_show | 27 | 0,26% |
| existing | 27 | 0,26% |
| methods | 26 | 0,25% |
| proposed | 25 | 0,24% |
| adversarial | 25 | 0,24% |
| history | 24 | 0,23% |
| submission | 23 | 0,22% |
| training | 23 | 0,22% |
| research | 22 | 0,21% |
Индексация Готовность: 10%
Индексирование
Есть ошибки
Разрешено ли индексирование страницы
?
Проверяет, не закрыта ли страница от индексации через robots.txt, meta robots или X-Robots-Tag. Страница, закрытая от индексации, не появится в поисковой выдаче.
Анкоров на странице 255 слишком много. Проведите ревизию и оптимизацию ссылок сайта.
Robots.txt
Найден корректный robots.txt
Файл управления сканированием сайта роботами
?
Robots.txt указывает поисковым роботам, какие страницы сканировать, а какие — нет. Ошибки в файле могут случайно закрыть важные разделы от индексации.
Robots.txt настроен корректно. Размер файла: 228998 байт. Загружен за: 1сек.
Проверяемая страница не запрещена в robots.txt.
Robots.txt доступен по постоянному адресу
Цепочка редиректов для файла robots.txt:
http://khoadoan.me/robots.txt
302 Found
https://mail-research.com/
200 OK
Показать содержимое robots.txt
<!DOCTYPE html>
<html>
<head>
<meta charset="utf-8">
<meta name="viewport" content="width=device-width, initial-scale=1, shrink-to-fit=no">
<meta http-equiv="X-UA-Compatible" content="IE=edge">
<title>
Khoa D Doan
</title>
<meta name="description" content="A simple, whitespace theme for academics. Based on [*folio](https://github.com/bogoli/-folio) design.
">
<!-- Open Graph -->
<!-- Bootstrap & MDB -->
<link href="https://stackpath.bootstrapcdn.com/bootstrap/4.5.2/css/bootstrap.min.css" rel="stylesheet" integrity="sha512-MoRNloxbStBcD8z3M/2BmnT+rg4IsMxPkXaGh2zD6LGNNFE80W3onsAhRcMAMrSoyWL9xD7Ert0men7vR8LUZg==" crossorigin="anonymous">
<link rel="stylesheet" href="https://cdnjs.cloudflare.com/ajax/libs/mdbootstrap/4.19.1/css/mdb.min.css" integrity="sha512-RO38pBRxYH3SoOprtPTD86JFOclM51/XTIdEPh5j8sj4tp8jmQIx26twG52UaLi//hQldfrh7e51WzP9wuP32Q==" crossorigin="anonymous" />
<link rel="stylesheet" href="https://cdn.jsdelivr.net/npm/bootstrap-icons@1.10.5/font/bootstrap-icons.css">
<!-- Fonts & Icons -->
<link rel="stylesheet" href="https://cdnjs.cloudflare.com/ajax/libs/font-awesome/5.14.0/css/all.min.css" integrity="sha512-1PKOgIY59xJ8Co8+NE6FZ+LOAZKjy+KY8iq0G4B3CyeY6wYHN3yt9PW0XpSriVlkMXe40PTKnXrLnZ9+fkDaog==" crossorigin="anonymous">
<link rel="stylesheet" href="https://cdnjs.cloudflare.com/ajax/libs/academicons/1.9.0/css/academicons.min.css" integrity="sha512-W4yqoT1+8NLkinBLBZko+dFB2ZbHsYLDdr50VElllRcNt2Q4/GSs6u71UHKxB7S6JEMCp5Ve4xjh3eGQl/HRvg==" crossorigin="anonymous">
<link rel="stylesheet" type="text/css" href="https://fonts.googleapis.com/css?family=Roboto:300,400,500,700|Roboto+Slab:100,300,400,500,700|Material+Icons">
<link rel="stylesheet" href="assets/font-awesome-4.7.0/css/font-awesome.min.css">
<!-- Code Syntax Highlighting -->
<!-- Very long loading using this file, so we swap -->
<!-- <link rel="stylesheet" href="https://gitcdn.link/repo/jwarby/jekyll-pygments-themes/master/github.css" /> -->
<link rel="stylesheet" href="https://combinatronics.io/jwarby/jekyll-pygments-themes/master/github.css" />
<!-- Styles -->
<!--
<link rel="icon" href="data:image/svg+xml,<svg xmlns=%22http://www.w3.org/2000/svg%22 viewBox=%220 0 100 100%22><text y=%22.9em%22 font-size=%2290%22></text></svg>">
-->
<link rel="icon" href="assets/img/icon.png">
<link rel="stylesheet" href="/assets/css/main.css">
<link rel="canonical" href="/">
<!-- JQuery -->
<!-- jQuery -->
<script src="https://cdnjs.cloudflare.com/ajax/libs/jquery/3.5.1/jquery.min.js" integrity="sha512-bLT0Qm9VnAYZDflyKcBaQ2gg0hSYNQrJ8RilYldYQ1FxQYoCLtUjuuRuZo+fjqhx/qtq/1itJ0C2ejDxltZVFg==" crossorigin="anonymous"></script>
<!-- Theming-->
<!-- Global site tag (gtag.js) - Google Analytics -->
<script async src="https://www.googletagmanager.com/gtag/js?id=G-5E8HDGL2HV"></script>
<script>
window.dataLayer = window.dataLayer || [];
function gtag() { dataLayer.push(arguments); }
gtag('js', new Date());
gtag('config', 'G-5E8HDGL2HV');
</script>
<script async
src="https://platform.x.com/widgets.js"
charset="utf-8">
</script>
<!-- MathJax -->
<script type="text/javascript">
window.MathJax = {
tex: {
tags: 'ams'
}
};
</script>
<script defer type="text/javascript" id="MathJax-script" src="https://cdn.jsdelivr.net/npm/mathjax@3.2.0/es5/tex-mml-chtml.js"></script>
<script defer src="https://polyfill.io/v3/polyfill.min.js?features=es6"></script>
</head>
<body class="fixed-top-nav ">
<!-- Header -->
<header>
<!-- Nav Bar -->
<nav id="navbar" class="navbar navbar-light navbar-expand-sm fixed-top">
<div class="container">
<div class="navbar-brand social">
<div class="logo">
<a href="https://mail-research.com"><img width="" height="30px" alt="blank" src="/assets/img/logo/logo-mail-v2.png"></a>
</div>
<!-- <div class="logo-mobile">
<a href=""><img width="32px" height="20px" alt="blank" src="/assets/img/logo/logo-mail-v2.png" /></a>
</div> -->
</div>
<!-- Navbar Toggle -->
<button class="navbar-toggler collapsed ml-auto" type="button" data-toggle="collapse" data-target="#navbarNav" aria-controls="navbarNav" aria-expanded="false" aria-label="Toggle navigation">
<span class="sr-only">Toggle navigation</span>
<span class="icon-bar top-bar"></span>
<span class="icon-bar middle-bar"></span>
<span class="icon-bar bottom-bar"></span>
</button>
<div class="collapse navbar-collapse text-right" id="navbarNav">
<ul class="navbar-nav ml-auto flex-nowrap">
<!-- About -->
<!-- <li class="nav-item ">
<a class="nav-link" href="/">
about
</a>
</li> -->
<!-- Other pages -->
<li class="nav-item ">
<a class="nav-link" href="/research">
research
</a>
</li>
<li class="nav-item ">
<a class="nav-link" href="/publications/">
publications
</a>
</li>
<li class="nav-item ">
<a class="nav-link" href="/team">
group
</a>
</li>
<li class="nav-item ">
<a class="nav-link" href="/photos/">
photos
</a>
</li>
<li class="nav-item ">
<a class="nav-link" href="/teaching/">
teaching
</a>
</li>
<li class="nav-item ">
<a class="nav-link" href="/contact">
contact
</a>
</li>
<!-- <li class="nav-item">
<a class="nav-link" href="/assets/pdf/khoadoan_cv_current.pdf">
CV
</a>
</li> -->
</ul>
</div>
</div>
</nav>
</header>
<!-- Content -->
<div class="container mt-5">
<div class="post">
<!-- <header class="post-header">
<h2 class="post-title"><span class="font-weight-bold">Machine Learning Algorithms</span> that Make Sense</h2>
<p class="desc">in constrained and large-scale settings with applications in <strong>Advertising</strong>, <strong>Healthcare</strong>, <strong>Sustainability</strong> (Climate, Computing, Agricultural), <strong>Social Goods</strong>...</p>
</header> -->
<article>
<div class="profile float-right">
<img class="img-fluid z-depth-1 rounded" src="/assets/img/team/khoa-d-doan-2.jpg">
<!-- <div class="address"> -->
<!-- <p>khoadoan at vt dot edu</p>
-->
<!-- </div> -->
<!-- <div class="team team-summary-large"> -->
<!-- </div> -->
<div class="social-small">
<div class="social contact-icons text-center">
<a href="mailto:%6B%68%6F%61%64%6F%61%6E%31%30%36@%67%6D%61%69%6C.%63%6F%6D"><i class="fas fa-envelope"></i></a>
<a href="https://github.com/mail-research" target="_blank" title="GitHub"><i class="fab fa-github" style="color: red;"></i></a>
<a href="https://www.linkedin.com/in/doankhoadang" target="_blank" title="LinkedIn"><i class="fab fa-linkedin"></i></a>
<a href="https://twitter.com/khoaddoan" target="_blank" title="Twitter"><i class="fab fa-twitter"></i></a>
<a href="https://facebook.com/mail.research" target="_blank" title="Twitter"><i class="fab fa-facebook"></i></a>
<!-- <break /> -->
<!-- <a href="https://github.com/mail-research" target="_blank" title="GitHub"><i class="fab fa-github" style="color:red"></i></a> -->
</div>
<!-- <div class="contact-note">khoadoan106 AT gmail.com
</div> -->
</div>
</div>
<!-- <h2 class="post-title"><span class="font-weight-bold">Machine Learning Algorithms</span> that Make Sense</h2>
<p class="desc">in constrained and large-scale settings with applications in <strong>Advertising</strong>, <strong>Healthcare</strong>, <strong>Sustainability</strong> (Climate, Computing, Agricultural), <strong>Social Goods</strong>...</p> -->
<!-- TITLE BLOCK -->
<!-- <div class="team-meta-standalone">
<h3 class="team-name">Khoa D Doan</h3>
<p class="team-title">Assistant Professor -- <a href='https://vinuni.edu.vn/college-of-engineering-computer-science/'>College of Engineering & Computer Science</a>, <a href='https://vinuni.edu.vn/'>VinUniversity</a></p>
<p class="team-title">Associate Director -- <a href='https://smarthealth.vinuni.edu.vn/'>VinUni-Illinois Smart Health Center</a></p>
[<a href="team">more about me and our group</a> | <a href="#open-office-hour-calendar">open office hour</a>
<i class="fa fa-calendar" aria-hidden="true"></i>
]
</div>
<hr /> -->
<div class="clearfix">
<p><span style="text-transform: uppercase; font-weight: bold">MAIL Research</span> aims to develop <span class="font-weight-bold">Machine Learning Algorithms</span> that <strong>Make Sense</strong> in <em>constrained</em> and <em>large-scale</em>, <em>societal</em> applications in <strong>Advertising</strong>, <strong>Healthcare</strong>, <strong>Sustainability</strong> (Remote Sensing, Computing, Agricultural)…
<!--
[[more about our research](research)] [[SAIL Research](https://sail-research.com/) <i class="fa fa-external-link" aria-hidden="true"></i>] [[photos](/photos) <i class="fa fa-picture-o" />] [<a href="#open-office-hour-calendar">open office hour</a>
<i class="fa fa-calendar" aria-hidden="true"></i>] --></p>
<p>[<a href="research">more about our research</a>] [<a href="https://sail-research.com/">SAIL Research</a> <i class="fa fa-external-link" aria-hidden="true"></i>] [<a href="#open-office-hour-calendar">open office hour</a>
<i class="fa fa-calendar" aria-hidden="true"></i>]</p>
<!-- **VinUni-Illinois Smart Health Center** ([VISHC](smarthealth.vinuni.edu.vn)) -- VISHC is open to collaborate with all researchers and industry practioners inside/oustide of Vietnam to solve healthcare related challenges with translational and innovative research, **together**. Please reach out via [[email](mailto:khoa.dd@vinuni.edu.vn)] for collaboration. [[more about VISHC](research/#VISHC)]. -->
<!-- **Center for Environmental Intelligence** ([CEI](https://cei.vinuni.edu.vn/)) -- *MAIL-Research is a member of CEI*. Please reach out via [[email](mailto:khoa.dd@vinuni.edu.vn)] for collaboration on *environmental monitoring*. [[more about CEI](research/#CEI)]. -->
<p><strong>Selected Press Coverage</strong>: <a href="https://khoahocphattrien.vn/cong-nghe/bill-melinda-gates-foundation-tai-tro-cho-vinuni-xay-dung-chatbot-cham-soc-suc-khoe-phu-khoa/20230814063044566p1c859.htm">khoahocphattrien</a>
, <a href="https://thanhnien.vn/du-an-khoa-hoc-viet-nam-duoc-to-chuc-quoc-te-tai-tro-1-trieu-bang-anh-185240608100441358.htm">Thanh Nien</a>, <a href="https://vnexpress.net/nha-khoa-hoc-viet-gianh-giai-thuong-toan-cau-ve-suc-khoe-4756007.html">VnExpress</a>, <a href="https://baodautu.vn/vinuni-gianh-giai-thuong-lon-nhat-tri-gia-1-trieu-bang-anh-trong-cuoc-thi-thu-thach-toan-cau-trinity-d217179.html">BaoDauTu</a>, <a href="https://dantri.com.vn/suc-khoe/vinuni-gianh-giai-thuong-cao-nhat-trong-cuoc-thi-thu-thach-trinity-20240608160801634.htm">DanTri</a>, <a href="https://www.vietnam.vn/vinuni-gianh-giai-thuong-lon-trong-cuoc-thi-toan-cau/">Vietnam.vn</a>, <a href="https://www.vietnam.vn/vinuni-gianh-giai-thuong-lon-trong-cuoc-thi-toan-cau/">Vietnam.vn</a>, <a href="https://sg.finance.yahoo.com/news/vinuni-won-1-million-global-104000008.html">Yahoo Finance</a>, <a href="https://www.benzinga.com/content/39240305/vinuni-won-1-million-global-grand-prize-in-the-trinity-challenge">Benzinga</a>, <a href="https://www.macaubusiness.com/vinuni-won-1-million-global-grand-prize-in-the-trinity-challenge/">Macau Business</a>, <a href="https://www.taiwannews.com.tw/news/5886511">Taiwan News</a>, <a href="https://technode.global/2024/06/25/from-science-fiction-to-reality-how-a-vietnamese-ai-platform-powered-by-human-ingenuity-could-save-millions-of-lives/">TNGlobal</a>, <a href="https://www.vingroup.net/tin-tuc-su-kien/bai-viet/3124/vinuni-gianh-giai-thuong-lon-nhat-tri-gia-1-trieu-bang-anh-trong-cuoc-thi-thu-thach-toan-cau-trinity">VinGroup</a>…</p>
<!-- Here at **MAIL**, We develop computational frameworks that enable existing complex/deep models to be more suitable for practical uses. We focus on improving the following aspects of existing models: (i) training/inference, (ii) realistic assumptions, (iii) algorithmic robustness, and (iv) efficiency in constrained settings. Most of our ML/AI solutions center around large-scale approaches that have low computational complexity and require less human effort.
[more about our research](research/) -->
<!-- <div class="team-meta-standalone">
<p class="team-name" style="font-size: 0.9rem;">Khoa D Doan</p>
<p class="team-title" style="font-size: 0.8rem;">Assistant Professor -- <a href='https://vinuni.edu.vn/college-of-engineering-computer-science/'>College of Engineering & Computer Science</a>, <a href='https://vinuni.edu.vn/'>VinUniversity</a></p>
<p class="team-title" style="font-size: 0.8rem;">Associate Director -- <a href='https://smarthealth.vinuni.edu.vn/'>VinUni-Illinois Smart Health Center</a></p>
</div> -->
</div>
<hr>
<div class="container pr-1 pl-0">
<div class="row justify-content-start">
<div class="col-12 col-md-5 col-lg-5 order-1 order-md-1" align="left">
<!-- <div class='jekyll-twitter-plugin'><p>There was a 'Moved Permanently' error fetching URL: 'https://x.com/khoaddoan'</p></div> -->
<div class="twitter-container">
<a class="twitter-timeline" data-width="400" data-height="250" data-dnt="true" href="https://x.com/khoaddoan">
Posts by @khoaddoan
</a>
</div>
</div>
<div class="annoucement col-12 col-md-7 col-lg-7 order-2 order-md-2">
<strong>We're open to research/industry collaborations in ML/CV/NLP</strong>:
<ul>
<!-- <li><strong>Intersted in Ph.D. position at VinUni?</strong> Please <a href="https://vinuni.edu.vn/phd/">apply</a> through the CS department and include my name as a potential advisor in your application.</li> -->
<!-- <li><strong>Interested in ML research as Research Assistant, Intern, PhD, Postdoc?</strong> Please fill in the form <a href="https://forms.gle/EXSe6Sb7SRsh6n2y6">here</a> (w. your CV, transcript, level of commitment, & description of what types of projects you want to work on).
<i>Only shortlisted candidates will be contacted!</i></li> -->
<li> I'm looking for (1-2) Postdocs working on responsible (safe/secure) ML, LLM Reasoning, mental-health NLP, please also fill in the form <a href="https://forms.gle/HQvvXUTSh1WfPoiS8">here</a>.
<strong>Deadline: Until Filled!</strong>
</li>
<!-- <li><strong style="color:red;font-style: italic;">:fire:Immediate Positions:fire:</strong>: Postdoc (1-2),
working on responsible (safe/secure) ML and mental-health NLP (co-advised with either
(a) <a href="https://scholar.google.com/citations?hl=en&user=hDLBEhkAAAAJ">Nitesh Chawla</a>, or (b)
<a href="https://scholar.google.com/citations?user=z7GCqT4AAAAJ&hl=en">Heng Ji</a> and
<a href="https://scholar.google.com/citations?user=GMcL_9kAAAAJ&hl=en">Dilek Hakkani-Tür</a>,
and VinUni Faculty), please also fill in the form <a href="https://forms.gle/HQvvXUTSh1WfPoiS8">here</a>.
<strong>Deadline: Until Filled!</strong>
</li> -->
<li><strong style="color:red;font-style: italic;">If you're interested in joining my team, please also fill in the form <a href="https://forms.gle/1VGzXMeJDnVq77FZA">here</a>.
</strong></li>
<!-- <li><strong style="color:red;font-style: italic;">:fire:Immediate Positions:fire:</strong>: Master by Research (2), PhD (2), Research Assistants (3-4)
also working on LLM/RL Fine-tuning, responsible (safe/secure) ML, mental-health NLP, and ML for Science (co-advised with either
(A) <a href="https://scholar.google.com/citations?hl=en&user=hDLBEhkAAAAJ">Nitesh Chawla</a> or (B)
<a href="https://scholar.google.com/citations?user=z7GCqT4AAAAJ&hl=en">Heng Ji</a>/<a href="https://scholar.google.com/citations?user=GMcL_9kAAAAJ&hl=en">Dilek Hakkani-Tür</a>,
or (C) <a href="https://scholar.google.com/citations?user=hKfga9oAAAAJ&hl=en">Chee Seng Chan</a>/<a href="https://khhuang.me/">Kuan-Hao Huang</a>
and
VinUni Faculty), please also fill in the form <a href="https://forms.gle/1VGzXMeJDnVq77FZA">here</a>.
<strong>Deadline: December 31st (positions are filling up quickly)!</strong>
</li> -->
<!-- <a href="https://cee.illinois.edu/directory/profile/thn">Thanh Huong Nguyen</a> -->
<!-- <li><strong style="color:red;font-style: italic;">:fire:Immediate Positions:fire:</strong>: Research Assistant (2-3),
UIUC or VinUni PhD/Postdoc (co-advised with <a href="https://blender.cs.illinois.edu/hengji.html">Heng Ji</a>,
<a href="https://minhdo.ece.illinois.edu/">Minh Do</a>,
<a href="https://aisecure.github.io/">Bo Li</a>, etc...), please also fill in the form <a href="https://forms.gle/EXSe6Sb7SRsh6n2y6">here</a>.
</li> -->
<!-- <li><strong style="color:red;font-style: italic;">:fire:Immediate Positions:fire:</strong>: Research Assistant (2),
working on low-resource/speech NLP, mental-health NLP, generative models (co-advised with UIUC, MBZUAI, and VinUni Faculty), please also fill in the form <a href="https://forms.gle/EXSe6Sb7SRsh6n2y6">here</a>.
</li> -->
<!-- <li><strong style="color:red;font-style: italic;">:fire:Immediate Positions:fire:</strong>: Software Engineer (3),
working on chatbot/LLM Application in Farm2Vets (our Trinity Challenge/MIT Solve Winning Solution).
</li> -->
<!-- <li><strong style="color:red;font-style: italic;">:fire:Several Positions (PhD, Master, Research Assistants, Software Developers):fire:</strong>
are SOON posted (by Mid November, 2024). Please check back in a few weeks for potential collaborations.
</li> -->
<li>
<strong>Other collaboration?</strong> Please reach out via <a href="mailto:khoadoan106@gmail.com">email</a>.</li>
</ul>
</div>
<!-- <div class="annoucement col-12 col-md-6 col-lg-6 order-1 order-md-1">
<strong>Prospective Ph.D. students:</strong>
<ul>
<li><strong>Not at VinUni?</strong> Please apply through the CS department and include my name as a potential advisor in your application.</li>
<li><strong>VinUni student?</strong> Please email me your CV, prior experiences, and how you would like to get involved.</li>
</ul>
<strong>If you are a VinUni undergraduate student:</strong>
<ul>
<li>Interested in engaging with ML research? Please reach out via email with your CV, transcript, level of commitment (e.g., 10-20 hours/week), and description of what types of projects you want to work on.</li>
</ul>
<strong>Prospective Visiting Students:</strong>
<ul>
<li>For short-term visiting, please fill in the form <a href="https://docs.google.com/forms/d/e/1FAIpQLSf7O3-3HmcNlZMQM5OnBZP0jygQMIXVM0ooGQulo6KSCN2dpg/viewform">here</a>. We will contact you if there is a match.</li>
</ul>
</div> -->
</div>
</div>
<!-- <hr /> -->
<!-- <div class="container pr-1 pl-0">
<div class="row justify-content-start">
<p>
<strong><a href="#open-office-hour-calendar">Open Office Hour</a></strong>- I will ocassionally be holding group open office hours for *anyone*.
Feel free to <a href="#open-office-hour-calendar">sign up</a> to connect, chat, or ask any questions.
</p>
<p>When I was a student, I had many questions that I had no ideas how to get help.
As a Vietnamese who went abroad for higher education,
and has worked in the industry and academia, and in applied projects as well as research projects,
I hope that, via this modest effort, I can share some experience with you, as well as addressing some questions you may have.
This effort is inspired by <a href="https://mlcollective.org/services/#mloh">ML Collective</a>.
</p>
</div>
</div> -->
<hr>
<div class="news">
<h2>news</h2>
<div class="table-responsive">
<table class="table table-sm table-borderless">
<tr>
<!-- <th>[07/08/2026]</th> -->
<th style="padding: 0px 10px 0px 0px;">[07/2026]</th>
<td style="padding: 0px 13px 0px 0px;">
<em>Accepted papers</em> - <a href="https://arxiv.org/abs/2605.17026">[COLM’26-a]</a> on data problem (forks in the road) of reasoning model’s coverage, <a href="https://arxiv.org/abs/2604.05971">[COLM’26-b]</a> on analyzing and mitigating center bias of CLIP-based VLMs.
</td>
</tr>
<tr>
<!-- <th>[06/04/2026]</th> -->
<th style="padding: 0px 10px 0px 0px;">[06/2026]</th>
<td style="padding: 0px 13px 0px 0px;">
<em>Accepted paper</em> - <a href="#">[Interspeech’26]</a> on a Vietnamese Speech Model with amazing performance!
</td>
</tr>
<tr>
<!-- <th>[05/01/2026]</th> -->
<th style="padding: 0px 10px 0px 0px;">[05/2026]</th>
<td style="padding: 0px 13px 0px 0px;">
<em>Accepted papers</em> - <a href="#">[ICML’26-a]</a> on token-level DPO, <a href="#">[ICML’26-b]</a> on data-efficient VLA adaptation, and <a href="#">[TMLR/ICML’26]</a> on continual learning.
</td>
</tr>
<tr>
<!-- <th>[03/15/2026]</th> -->
<th style="padding: 0px 10px 0px 0px;">[03/2026]</th>
<td style="padding: 0px 13px 0px 0px;">
<img class="emoji" title=":tada:" alt=":tada:" src="https://github.githubassets.com/images/icons/emoji/unicode/1f389.png" height="20" width="20"> Invited to join the <strong>NeurIPS 2026’s Organizing Team</strong> as <strong>Workshop Chair</strong>.
</td>
</tr>
<tr>
<!-- <th>[03/05/2026]</th> -->
<th style="padding: 0px 10px 0px 0px;">[03/2026]</th>
<td style="padding: 0px 13px 0px 0px;">
<img class="emoji" title=":fire:" alt=":fire:" src="https://github.githubassets.com/images/icons/emoji/unicode/1f525.png" height="20" width="20"> Our <a href="https://tridentatmm26mgc.github.io/trident.github.io/">TRIDENT challenge</a> (w. <a href="https://www.csie.ntu.edu.tw/~wenhuang/">Wen-Huang Cheng</a>, <a href="http://basiclab.lab.nycu.edu.tw/">Hong-Han Shuai</a>, <a href="http://www.hongxiaxie.net/">Hongxia Xie</a>, <a href="https://lynn-ling-lo.github.io/">Ling Lo</a>, <a href="https://github.com/j1anglin">Jian-Yu Jiang-Lin</a>, <a href="https://alberthkyhky.github.io/">Kang-Yang Huang</a>, and <a href="https://www.linkedin.com/in/ppofzl">Ling Zou</a>) on multimodal deepfake perception, detection, and hallucination accepted at ACM MM’26.
</td>
</tr>
<tr>
<!-- <th>[02/20/2026]</th> -->
<th style="padding: 0px 10px 0px 0px;">[02/2026]</th>
<td style="padding: 0px 13px 0px 0px;">
<em>Accepted papers</em> - <a href="#">[CVPR’26]</a> on federarated domain generalization and <a href="#">[CVPR’26-Findings]</a> personalized FL with hypernetworks.
</td>
</tr>
<tr>
<!-- <th>[01/14/2026]</th> -->
<th style="padding: 0px 10px 0px 0px;">[01/2026]</th>
<td style="padding: 0px 13px 0px 0px;">
<em>Accepted papers</em> - <a href="https://openreview.net/forum?id=9NnhVME4Q6">[TMLR’26]</a> (<strong>Featured</strong>) on a new continual learning theme, <a href="#">[EACL’26]</a> on LLM watermarking (w. minimal text equality degradation), and <a href="https://openreview.net/forum?id=cpwbXHvd2h">[ICLR’26]</a> on 2D-3D fusion for predicting molecular properties.
</td>
</tr>
<tr>
<!-- <th>[11/15/2025]</th> -->
<th style="padding: 0px 10px 0px 0px;">[11/2025]</th>
<td style="padding: 0px 13px 0px 0px;">
Will serve as Invited <strong>Area Chair</strong> for <strong>ICML</strong> 2026.
</td>
</tr>
<tr>
<!-- <th>[11/08/2025]</th> -->
<th style="padding: 0px 10px 0px 0px;">[11/2025]</th>
<td style="padding: 0px 13px 0px 0px;">
Gave a talk at National Taiwan University on novel problem solving and LLMs.
</td>
</tr>
<tr>
<!-- <th>[11/01/2025]</th> -->
<th style="padding: 0px 10px 0px 0px;">[11/2025]</th>
<td style="padding: 0px 13px 0px 0px;">
<em>Accepted paper</em> - <a href="">[AAAI’26-a]</a> on clean backdoor attacks with data distillation.
</td>
</tr>
<tr>
<!-- <th>[09/01/2025]</th> -->
<th style="padding: 0px 10px 0px 0px;">[09/2025]</th>
<td style="padding: 0px 13px 0px 0px;">
<em>Accepted papers</em> - <a href="https://openreview.net/forum?id=ltPRj2nthL">[NeurIPS’25-a]</a> on fixing DPO/IPO’s overfitting, <a href="https://openreview.net/forum?id=dVIx32Lq7J">[NeurIPS’25-b]</a> on parameter attributions in and inference control fo diffusion models, <a href="https://openreview.net/forum?id=cFVQJepi4e">[NeurIPS’25-c]</a> efficient inference with token merging in 3D Point Cloud, and <a href="https://openreview.net/forum?id=FwOmMo8ek9">[NeurIPS-W’25-d]</a> on clean backdoor attacks with data distillation.
</td>
</tr>
<tr>
<!-- <th>[08/21/2025]</th> -->
<th style="padding: 0px 10px 0px 0px;">[08/2025]</th>
<td style="padding: 0px 13px 0px 0px;">
Will serve as Invited <strong>Area Chair</strong> for <strong>AISTATS</strong> and <strong>ICLR</strong> 2026.
</td>
</tr>
<tr>
<!-- <th>[07/25/2025]</th> -->
<th style="padding: 0px 10px 0px 0px;">[07/2025]</th>
<td style="padding: 0px 13px 0px 0px;">
<img class="emoji" title=":fire:" alt=":fire:" src="https://github.githubassets.com/images/icons/emoji/unicode/1f525.png" height="20" width="20"> Will serve on the Editorial Board of <a href="https://dl.acm.org/journal/ailet">ACM AI Letters</a>, a premier venue focusing on impactful, concise, and timely communications in AI.
</td>
</tr>
<tr>
<!-- <th>[07/01/2025]</th> -->
<th style="padding: 0px 10px 0px 0px;">[07/2025]</th>
<td style="padding: 0px 13px 0px 0px;">
<div class="d-inline-block" tabindex="0" data-toggle="tooltip" title="Majority of work done by MAIL/SAIL members!">
<i class="bi bi-rocket-takeoff-fill"></i>
</div>
Congrats several MAIL members for Grad Acceptances: <a href="team/#Chau%20Hoai%20Tran">Chau H Tran</a> and <a href="team/">Nguyen K Hoang</a> and <a href="https://mtuann.github.io/">Tuan M Nguyen</a> → UIUC, <a href="team/#Hieu%20Ngoc%20Nguyen">Hieu N Nguyen</a> → Penn State, <a href="team/#Quan%20Huu%20Do">Quan H Do</a> → Institut Polytechnique de Paris
</td>
</tr>
<tr>
<!-- <th>[06/01/2025]</th> -->
<th style="padding: 0px 10px 0px 0px;">[06/2025]</th>
<td style="padding: 0px 13px 0px 0px;">
<em>Accepted papers</em> - <a href="https://arxiv.org/pdf/2504.10415">[ICML’25-a]</a> (<strong>Oral</strong>) on evaluating novel equation discovery of LLM-based symbolic regression methods and <a href="https://openreview.net/pdf?id=UlW8UhEiei">[ICML-W’25-b]</a> on fragment-aware, structure-guided graph transformer.
</td>
</tr>
</table>
</div>
</div>
<hr>
<div class="publications">
<h2>selected publications [<a href="publications">full list</a>]</h2>
<div class="d-inline-block" tabindex="0" data-toggle="tooltip" title="Majority of work done or significant contribution by MAIL/SAIL members!">
<abbr class="badge bg-danger"><i class="bi bi-check2"></i></abbr> <i>Majority of work done or significant contribution by MAIL/SAIL members!!</i>
</div>
<br>
<br>
<ol class="bibliography">
<li>
<div class="row">
<!-- <div class="col-sm-1 abbr">
<div>
<abbr class="badge">COLM</abbr>
</div>
</div> -->
<div id="nguyen2026forks" class="col-sm-12">
<div class="title">
<span class="abbr">
<abbr class="badge">COLM</abbr>
</span>
<span class="d-inline-block" tabindex="0" data-toggle="tooltip" title="Majority of work done or significant contribution by MAIL/SAIL members!">
<abbr class="badge bg-danger"><i class="bi bi-check2"></i></abbr>
</span>
Why Do Reasoning Models Lose Coverage? The Role of Data and Forks in the Road
</div>
<div class="author">
<span style="font-weight: bold;">Ngoc-Hieu Nguyen</span>,
Parshin Shojaee,
<span style="font-weight: bold;">Phuc M Nguyen</span>,
Nan Zhang,
<a href="https://people.cs.vt.edu/reddy/index.html" target="_blank">Chandan K Reddy</a>,
<u>Khoa D Doan</u>,
and Rui Zhang
</div>
<div class="periodical">
<em>Conference on Language Modeling</em>
2026
</div>
<div class="links">
<a class="abstract btn btn-sm z-depth-0" role="button">Abs</a>
<a class="bibtex btn btn-sm z-depth-0" role="button">Bib</a>
<a href="https://arxiv.org/pdf/2605.17026" class="btn btn-sm z-depth-0" role="button" target="_blank">PDF</a>
<a href="https://github.com/psunlpgroup/reasoning_forks" class="btn btn-sm z-depth-0" role="button" target="_blank">Code</a>
<a class="submissions btn btn-sm z-depth-0" role="button">Submission History</a>
</div>
<!-- Hidden abstract block -->
<div class="abstract hidden">
<p>Recent progress in large language models has led to the emergence of reasoning models, which have shown strong performance on complex tasks through specialized fine-tuning procedures. While these methods reliably improve pass@1 accuracy, prior works have observed that they show a coverage shrinkage behavior, where pass@k degrades relative to the base model. In this paper, we investigate the reasoning shrinkage arise under SFT-based post-training. We hypothesize that this behavior is driven by properties of the fine-tuning data, specifically related to decision points or "forks in the road" scenarios where model faces indecipherable patterns with multiple valid reasoning paths. To test this hypothesis, we design controlled case studies that simulate such decision-point settings, spanning indecipherable nodes in graph branching, and reasoning modes. By tracking post-training dynamics in these settings, we find that the shrinkage phenomenon is tightly correlated with the prevalence of decision-point scenarios in the training data. We also demonstrate that this shrinkage behavior can be partially mitigated through targeted data synthesis design of decision-points, and a more systematic diversity-encouraging decoding mechanism. Our findings identify data-centric factors as a key driver of shrinkage in reasoning models and highlight diversity-aware designs as an effective lever for controlling it.</p>
</div>
<!-- Hidden bibtex block -->
<div class="bibtex hidden">
<figure class="highlight"><pre><code class="language-bibtex" data-lang="bibtex"><span class="nc">@article</span><span class="p">{</span><span class="nl">nguyen2026forks</span><span class="p">,</span>
<span class="na">title</span> <span class="p">=</span> <span class="s">{Why Do Reasoning Models Lose Coverage? The Role of Data and Forks in the Road}</span><span class="p">,</span>
<span class="na">author</span> <span class="p">=</span> <span class="s">{Nguyen, Ngoc-Hieu and Shojaee, Parshin and Nguyen, Phuc M and Zhang, Nan and Reddy, Chandan K and Doan, Khoa D and Zhang, Rui}</span><span class="p">,</span>
<span class="na">year</span> <span class="p">=</span> <span class="s">{2026}</span><span class="p">,</span>
<span class="na">bibtex_show</span> <span class="p">=</span> <span class="s">{true}</span><span class="p">,</span>
<span class="na">abbr</span> <span class="p">=</span> <span class="s">{COLM}</span><span class="p">,</span>
<span class="na">journal</span> <span class="p">=</span> <span class="s">{Conference on Language Modeling}</span><span class="p">,</span>
<span class="na">pdf</span> <span class="p">=</span> <span class="s">{https://arxiv.org/pdf/2605.17026}</span><span class="p">,</span>
<span class="na">code</span> <span class="p">=</span> <span class="s">{https://github.com/psunlpgroup/reasoning_forks}</span><span class="p">,</span>
<span class="na">selected</span> <span class="p">=</span> <span class="s">{true}</span><span class="p">,</span>
<span class="na">submissions</span> <span class="p">=</span> <span class="s">{COLM'26}</span>
<span class="p">}</span></code></pre></figure>
</div>
<!-- Hidden bibtex block -->
<div class="submissions hidden">
<p>COLM'26</p>
</div>
</div>
</div>
</li>
<li>
<div class="row">
<!-- <div class="col-sm-1 abbr">
<div>
<abbr class="badge">TMLR</abbr>
</div>
</div> -->
<div id="nguyen2026irl" class="col-sm-12">
<div class="title">
<span class="abbr">
<abbr class="badge">TMLR</abbr>
</span>
<span class="d-inline-block" tabindex="0" data-toggle="tooltip" title="Majority of work done or significant contribution by MAIL/SAIL members!">
<abbr class="badge bg-danger"><i class="bi bi-check2"></i></abbr>
</span>
<abbr class="badge" style="background-color: #04a96d;">FEATURED</abbr>
Retrospective Feature Estimation for Continual Learning
</div>
<div class="author">
<span style="font-weight: bold;">Nghia D Nguyen</span>,
<a href="https://hieunt91.github.io/" target="_blank">Hieu T Nguyen</a>,
<a href="https://angli.ai/" target="_blank">Ang Li</a>,
<span style="font-weight: bold;">Hoang V Pham</span>,
<a href="https://vietanhnguyen.net/" target="_blank">Viet Anh Nguyen</a>,
and <u>Khoa D Doan</u>
</div>
<div class="periodical">
<em>Transactions on Machine Learning Research</em>
2026
</div>
<div class="links">
<a class="abstract btn btn-sm z-depth-0" role="button">Abs</a>
<a class="bibtex btn btn-sm z-depth-0" role="button">Bib</a>
<a href="https://openreview.net/pdf?id=9NnhVME4Q6" class="btn btn-sm z-depth-0" role="button" target="_blank">PDF</a>
<a href="https://github.com/mail-research/retrospective-feature-estimation" class="btn btn-sm z-depth-0" role="button" target="_blank">Code</a>
<a class="submissions btn btn-sm z-depth-0" role="button">Submission History</a>
</div>
<!-- Hidden abstract block -->
<div class="abstract hidden">
<p>The intrinsic capability to continuously learn a changing data stream is a desideratum of deep neural networks (DNNs). However, current DNNs suffer from catastrophic forgetting, which interferes with remembering past knowledge. To mitigate this issue, existing Continual Learning (CL) approaches often retain exemplars for replay, regularize learning, or allocate dedicated capacity for new tasks. This paper investigates an unexplored direction for CL called Retrospective Feature Estimation (RFE). RFE learns to reverse feature changes by aligning the features from the current trained DNN backward to the feature space of the old task, where performing predictions is easier. This retrospective process utilizes a chain of small feature mapping networks called retrospector modules. Empirical experiments on several CL benchmarks, including CIFAR10, CIFAR100, and Tiny ImageNet, demonstrate the effectiveness and potential of this novel CL direction compared to existing representative CL methods, motivating further research into retrospective mechanisms as a principled alternative for mitigating catastrophic forgetting in CL. Code is available at: https://github.com/mail-research/retrospective-feature-estimation</p>
</div>
<!-- Hidden bibtex block -->
<div class="bibtex hidden">
<figure class="highlight"><pre><code class="language-bibtex" data-lang="bibtex"><span class="nc">@article</span><span class="p">{</span><span class="nl">nguyen2026irl</span><span class="p">,</span>
<span class="na">title</span> <span class="p">=</span> <span class="s">{Retrospective Feature Estimation for Continual Learning}</span><span class="p">,</span>
<span class="na">author</span> <span class="p">=</span> <span class="s">{Nguyen, Nghia D and Nguyen, Hieu T and Li, Ang and Pham, Hoang V and Nguyen, Viet Anh and Doan, Khoa D}</span><span class="p">,</span>
<span class="na">year</span> <span class="p">=</span> <span class="s">{2026}</span><span class="p">,</span>
<span class="na">bibtex_show</span> <span class="p">=</span> <span class="s">{true}</span><span class="p">,</span>
<span class="na">abbr</span> <span class="p">=</span> <span class="s">{TMLR}</span><span class="p">,</span>
<span class="na">journal</span> <span class="p">=</span> <span class="s">{Transactions on Machine Learning Research}</span><span class="p">,</span>
<span class="na">pdf</span> <span class="p">=</span> <span class="s">{https://openreview.net/pdf?id=9NnhVME4Q6}</span><span class="p">,</span>
<span class="na">code</span> <span class="p">=</span> <span class="s">{https://github.com/mail-research/retrospective-feature-estimation}</span><span class="p">,</span>
<span class="na">selected</span> <span class="p">=</span> <span class="s">{true}</span><span class="p">,</span>
<span class="na">note</span> <span class="p">=</span> <span class="s">{FEATURED}</span><span class="p">,</span>
<span class="na">submissions</span> <span class="p">=</span> <span class="s">{ICML'24 -- NeurIPS'24 -- ICLR'25 -- ICCV'25 -- WACV'26 -- TMLR'26}</span>
<span class="p">}</span></code></pre></figure>
</div>
<!-- Hidden bibtex block -->
<div class="submissions hidden">
<p>ICML'24 -- NeurIPS'24 -- ICLR'25 -- ICCV'25 -- WACV'26 -- TMLR'26</p>
</div>
</div>
</div>
</li>
<li>
<div class="row">
<!-- <div class="col-sm-1 abbr">
<div>
<abbr class="badge">EACL</abbr>
</div>
</div> -->
<div id="cao2026sparsellmwatermark" class="col-sm-12">
<div class="title">
<span class="abbr">
<abbr class="badge">EACL</abbr>
</span>
<span class="d-inline-block" tabindex="0" data-toggle="tooltip" title="Majority of work done or significant contribution by MAIL/SAIL members!">
<abbr class="badge bg-danger"><i class="bi bi-check2"></i></abbr>
</span>
SpARK: An Embarrassingly Simple Sparse Watermarking in LLMs with Enhanced Text Quality
</div>
<div class="author">
<span style="font-weight: bold;">Cao-Duy Hoang</span>,
<span style="font-weight: bold;">Hung T. Q. Le</span>,
Rui Chu,
<a href="http://research.baidu.com/People/index-view?id=111" target="_blank">Ping Li</a>,
<a href="https://www.cs.rit.edu/~wjz/" target="_blank">Weijie Zhao</a>,
<a href="http://ylao.people.clemson.edu/" target="_blank">Yingjie Lao</a>,
and <u>Khoa D Doan</u>
</div>
<div class="periodical">
<em>In Findings of the European Chapter of the Association for Computational Linguistics</em>
2026
</div>
<div class="links">
<a class="abstract btn btn-sm z-depth-0" role="button">Abs</a>
<a class="bibtex btn btn-sm z-depth-0" role="button">Bib</a>
<a href="https://github.com/mail-research/sparse-llm-watermarking" class="btn btn-sm z-depth-0" role="button" target="_blank">Code</a>
<a class="submissions btn btn-sm z-depth-0" role="button">Submission History</a>
</div>
<!-- Hidden abstract block -->
<div class="abstract hidden">
<p>With the widespread adoption of Large Language Models (LLMs), concerns about potential misuse have emerged. To this end, watermarking has been adapted to LLM, enabling a simple and effective way to detect and monitor generated text. However, while the existing methods can differentiate between watermarked and unwatermarked text with high accuracy, they often face a trade-off between the quality of the generated text and the effectiveness of the watermarking process. In this work, we present a novel type of LLM watermark, Sparse WatermARK (or SpARK), which aims to mitigate this trade-off by applying watermarks to a small subset of generated tokens distributed across the text. To demonstrate this type of watermark, we introduce two novel variants, SpARK-P and SpARK-R, which achieve sparsity by anchoring watermarked tokens to words that have specific Part-of-Speech (POS) tags and specific hash values w.r.t a pseudorandom hash function, respectively. Our experimental results demonstrate that the proposed watermarking schemes, albeit embarrassingly simple, are incredibly effective, achieving high detectability while generating text that outperforms previous LLM watermarking methods in quality across various tasks. SpARK further advances the watermarking capability for LLMs while maintaining their generated text quality.</p>
</div>
<!-- Hidden bibtex block -->
<div class="bibtex hidden">
<figure class="highlight"><pre><code class="language-bibtex" data-lang="bibtex"><span class="nc">@inproceedings</span><span class="p">{</span><span class="nl">cao2026sparsellmwatermark</span><span class="p">,</span>
<span class="na">title</span> <span class="p">=</span> <span class="s">{SpARK: An Embarrassingly Simple Sparse Watermarking in LLMs with Enhanced Text Quality}</span><span class="p">,</span>
<span class="na">author</span> <span class="p">=</span> <span class="s">{Hoang, Cao-Duy and Le, Hung T. Q. and Chu, Rui and Li, Ping and Zhao, Weijie and Lao, Yingjie and Doan, Khoa D}</span><span class="p">,</span>
<span class="na">year</span> <span class="p">=</span> <span class="s">{2026}</span><span class="p">,</span>
<span class="na">bibtex_show</span> <span class="p">=</span> <span class="s">{true}</span><span class="p">,</span>
<span class="na">abbr</span> <span class="p">=</span> <span class="s">{EACL}</span><span class="p">,</span>
<span class="na">booktitle</span> <span class="p">=</span> <span class="s">{Findings of the European Chapter of the Association for Computational Linguistics}</span><span class="p">,</span>
<span class="na">code</span> <span class="p">=</span> <span class="s">{https://github.com/mail-research/sparse-llm-watermarking}</span><span class="p">,</span>
<span class="na">selected</span> <span class="p">=</span> <span class="s">{true}</span><span class="p">,</span>
<span class="na">submissions</span> <span class="p">=</span> <span class="s">{NeurIPS'24 -- ICLR'25 -- COLM'25 -- ARR'05-25 -- ARR'10-25 (EACL)}</span>
<span class="p">}</span></code></pre></figure>
</div>
<!-- Hidden bibtex block -->
<div class="submissions hidden">
<p>NeurIPS'24 -- ICLR'25 -- COLM'25 -- ARR'05-25 -- ARR'10-25 (EACL)</p>
</div>
</div>
</div>
</li>
<li>
<div class="row">
<!-- <div class="col-sm-1 abbr">
<div>
<abbr class="badge">NeurIPS</abbr>
</div>
</div> -->
<div id="nguyen2025merging3d" class="col-sm-12">
<div class="title">
<span class="abbr">
<abbr class="badge">NeurIPS</abbr>
</span>
<span class="d-inline-block" tabindex="0" data-toggle="tooltip" title="Majority of work done or significant contribution by MAIL/SAIL members!">
<abbr class="badge bg-danger"><i class="bi bi-check2"></i></abbr>
</span>
How Many Tokens Do 3D Point Cloud Transformer Architectures Really Need?
</div>
<div class="author">
Tuan A Tran,
Duy MH Nguyen,
<span style="font-weight: bold;">Chau H Tran</span>,
and others
</div>
<div class="periodical">
<em>In Advances in Neural Information Processing Systems</em>
2025
</div>
<div class="links">
<a class="abstract btn btn-sm z-depth-0" role="button">Abs</a>
<a class="bibtex btn btn-sm z-depth-0" role="button">Bib</a>
<a class="submissions btn btn-sm z-depth-0" role="button">Submission History</a>
</div>
<!-- Hidden abstract block -->
<div class="abstract hidden">
<p>Recent advances in 3D point cloud transformers have led to state-of-the-art results in tasks such as semantic segmentation and reconstruction. However, these models typically rely on dense token representations, incurring high computational and memory costs during training and inference. In this work, we present an efficient token merging strategy that drastically reduces the token count by up to 90–95\% while preserving competitive performance. Our approach estimates token importance by leveraging spatial structures within the 3D point cloud, enabling aggressive token reduction with minimal degradation in accuracy. This finding challenges the prevailing assumption that more tokens inherently yield better performance and highlights that many current models are over-tokenized and under-optimized for scalability. We validate our method across multiple 3D vision tasks and show consistent improvements in computational efficiency. Our ongoing work will release code and detailed benchmarks to support reproducibility and further system-level exploration of efficient foundation models for 3D data.</p>
</div>
<!-- Hidden bibtex block -->
<div class="bibtex hidden">
<figure class="highlight"><pre><code class="language-bibtex" data-lang="bibtex"><span class="nc">@inproceedings</span><span class="p">{</span><span class="nl">nguyen2025merging3d</span><span class="p">,</span>
<span class="na">title</span> <span class="p">=</span> <span class="s">{How Many Tokens Do 3D Point Cloud Transformer Architectures Really Need?}</span><span class="p">,</span>
<span class="na">author</span> <span class="p">=</span> <span class="s">{Tran, Tuan A and Nguyen, Duy MH and Tran, Chau H and others}</span><span class="p">,</span>
<span class="na">booktitle</span> <span class="p">=</span> <span class="s">{Advances in Neural Information Processing Systems}</span><span class="p">,</span>
<span class="na">year</span> <span class="p">=</span> <span class="s">{2025}</span><span class="p">,</span>
<span class="na">bibtex_show</span> <span class="p">=</span> <span class="s">{true}</span><span class="p">,</span>
<span class="na">abbr</span> <span class="p">=</span> <span class="s">{NeurIPS}</span><span class="p">,</span>
<span class="na">selected</span> <span class="p">=</span> <span class="s">{true}</span><span class="p">,</span>
<span class="na">submissions</span> <span class="p">=</span> <span class="s">{NeurIPS'25}</span>
<span class="p">}</span></code></pre></figure>
</div>
<!-- Hidden bibtex block -->
<div class="submissions hidden">
<p>NeurIPS'25</p>
</div>
</div>
</div>
</li>
<li>
<div class="row">
<!-- <div class="col-sm-1 abbr">
<div>
<abbr class="badge">NeurIPS</abbr>
</div>
</div> -->
<div id="nguyen2025ais" class="col-sm-12">
<div class="title">
<span class="abbr">
<abbr class="badge">NeurIPS</abbr>
</span>
<span class="d-inline-block" tabindex="0" data-toggle="tooltip" title="Majority of work done or significant contribution by MAIL/SAIL members!">
<abbr class="badge bg-danger"><i class="bi bi-check2"></i></abbr>
</span>
Mitigating Reward Over-optimization in Direct Alignment Algorithms with Adaptive Importance Sampling
</div>
<div class="author">
<span style="font-weight: bold;">Phuc M Nguyen</span>,
<span style="font-weight: bold;">Ngoc-Hieu Nguyen</span>,
Binh T Nguyen,
and <u>Khoa D Doan</u>
</div>
<div class="periodical">
<em>In Advances in Neural Information Processing Systems</em>
2025
</div>
<div class="links">
<a class="abstract btn btn-sm z-depth-0" role="button">Abs</a>
<a class="bibtex btn btn-sm z-depth-0" role="button">Bib</a>
<a href="https://arxiv.org/abs/2506.08681" class="btn btn-sm z-depth-0" role="button" target="_blank">PDF</a>
<a href="https://github.com/mail-research/AIS-Sampling4DAAs" class="btn btn-sm z-depth-0" role="button" target="_blank">Code</a>
<a class="submissions btn btn-sm z-depth-0" role="button">Submission History</a>
</div>
<!-- Hidden abstract block -->
<div class="abstract hidden">
<p>Direct Alignment Algorithms (DAAs) such as Direct Preference Optimization (DPO) have emerged as alternatives to the standard Reinforcement Learning from Human Feedback (RLHF) for aligning large language models (LLMs) with human values. However, these methods are more susceptible to over-optimization, in which the model drifts away from the reference policy, leading to degraded performance as training progresses. This paper proposes a novel importance-sampling approach to mitigate the over-optimization problem of offline DAAs. This approach, called (IS-DAAs), multiplies the DAA objective with an importance ratio that accounts for the reference policy distribution. IS-DAAs additionally avoid the high variance issue associated with importance sampling by clipping the importance ratio to a maximum value. Our extensive experiments demonstrate that IS-DAAs can effectively mitigate over-optimization, especially under low regularization strength, and achieve better performance than other methods designed to address this problem.</p>
</div>
<!-- Hidden bibtex block -->
<div class="bibtex hidden">
<figure class="highlight"><pre><code class="language-bibtex" data-lang="bibtex"><span class="nc">@inproceedings</span><span class="p">{</span><span class="nl">nguyen2025ais</span><span class="p">,</span>
<span class="na">title</span> <span class="p">=</span> <span class="s">{Mitigating Reward Over-optimization in Direct Alignment Algorithms with Adaptive Importance Sampling}</span><span class="p">,</span>
<span class="na">author</span> <span class="p">=</span> <span class="s">{Nguyen, Phuc M and Nguyen, Ngoc-Hieu and Nguyen, Binh T and Doan, Khoa D}</span><span class="p">,</span>
<span class="na">booktitle</span> <span class="p">=</span> <span class="s">{Advances in Neural Information Processing Systems}</span><span class="p">,</span>
<span class="na">year</span> <span class="p">=</span> <span class="s">{2025}</span><span class="p">,</span>
<span class="na">bibtex_show</span> <span class="p">=</span> <span class="s">{true}</span><span class="p">,</span>
<span class="na">abbr</span> <span class="p">=</span> <span class="s">{NeurIPS}</span><span class="p">,</span>
<span class="na">pdf</span> <span class="p">=</span> <span class="s">{https://arxiv.org/abs/2506.08681}</span><span class="p">,</span>
<span class="na">selected</span> <span class="p">=</span> <span class="s">{true}</span><span class="p">,</span>
<span class="na">code</span> <span class="p">=</span> <span class="s">{https://github.com/mail-research/AIS-Sampling4DAAs}</span><span class="p">,</span>
<span class="na">submissions</span> <span class="p">=</span> <span class="s">{ICLR'25 -- COLM'25 (withdraw-missing title) -- NeurIPS'25}</span>
<span class="p">}</span></code></pre></figure>
</div>
<!-- Hidden bibtex block -->
<div class="submissions hidden">
<p>ICLR'25 -- COLM'25 (withdraw-missing title) -- NeurIPS'25</p>
</div>
</div>
</div>
</li>
<li>
<div class="row">
<!-- <div class="col-sm-1 abbr">
<div>
<abbr class="badge">NeurIPS</abbr>
</div>
</div> -->
<div id="nguyen2025cad" class="col-sm-12">
<div class="title">
<span class="abbr">
<abbr class="badge">NeurIPS</abbr>
</span>
<span class="d-inline-block" tabindex="0" data-toggle="tooltip" title="Majority of work done or significant contribution by MAIL/SAIL members!">
<abbr class="badge bg-danger"><i class="bi bi-check2"></i></abbr>
</span>
Unveiling Concept Attribution in Diffusion Models
</div>
<div class="author">
<span style="font-weight: bold;">Quang H Nguyen</span>,
<span style="font-weight: bold;">Phan Hoang</span>,
and <u>Khoa D Doan</u>
</div>
<div class="periodical">
<em>In Advances in Neural Information Processing Systems</em>
2025
</div>
<div class="links">
<a class="abstract btn btn-sm z-depth-0" role="button">Abs</a>
<a class="bibtex btn btn-sm z-depth-0" role="button">Bib</a>
<a href="https://arxiv.org/abs/2412.02542" class="btn btn-sm z-depth-0" role="button" target="_blank">PDF</a>
<a href="https://github.com/mail-research/CAD-attribution4diffusion" class="btn btn-sm z-depth-0" role="button" target="_blank">Code</a>
<a class="submissions btn btn-sm z-depth-0" role="button">Submission History</a>
</div>
<!-- Hidden abstract block -->
<div class="abstract hidden">
<p>Diffusion models have shown remarkable abilities in generating realistic and high-quality images from text prompts. However, a trained model remains black-box; little do we know about the role of its components in exhibiting a concept such as objects or styles. Recent works employ causal tracing to localize layers storing knowledge in generative models without showing how those layers contribute to the target concept. In this work, we approach the model interpretability problem from a more general perspective and pose a question: \textit{``How do model components work jointly to demonstrate knowledge?''}. We adapt component attribution to decompose diffusion models, unveiling how a component contributes to a concept. Our framework allows effective model editing, in particular, we can erase a concept from diffusion models by removing positive components while remaining knowledge of other concepts. Surprisingly, we also show there exist components that contribute negatively to a concept, which has not been discovered in the knowledge localization approach. Experimental results confirm the role of positive and negative components pinpointed by our framework, depicting a complete view of interpreting generative models. Our code is available at https://github.com/mail-research/CAD-attribution4diffusion</p>
</div>
<!-- Hidden bibtex block -->
<div class="bibtex hidden">
<figure class="highlight"><pre><code class="language-bibtex" data-lang="bibtex"><span class="nc">@inproceedings</span><span class="p">{</span><span class="nl">nguyen2025cad</span><span class="p">,</span>
<span class="na">title</span> <span class="p">=</span> <span class="s">{Unveiling Concept Attribution in Diffusion Models}</span><span class="p">,</span>
<span class="na">author</span> <span class="p">=</span> <span class="s">{Nguyen, Quang H and Hoang, Phan and Doan, Khoa D}</span><span class="p">,</span>
<span class="na">booktitle</span> <span class="p">=</span> <span class="s">{Advances in Neural Information Processing Systems}</span><span class="p">,</span>
<span class="na">year</span> <span class="p">=</span> <span class="s">{2025}</span><span class="p">,</span>
<span class="na">bibtex_show</span> <span class="p">=</span> <span class="s">{true}</span><span class="p">,</span>
<span class="na">abbr</span> <span class="p">=</span> <span class="s">{NeurIPS}</span><span class="p">,</span>
<span class="na">pdf</span> <span class="p">=</span> <span class="s">{https://arxiv.org/abs/2412.02542}</span><span class="p">,</span>
<span class="na">selected</span> <span class="p">=</span> <span class="s">{true}</span><span class="p">,</span>
<span class="na">code</span> <span class="p">=</span> <span class="s">{https://github.com/mail-research/CAD-attribution4diffusion}</span><span class="p">,</span>
<span class="na">submissions</span> <span class="p">=</span> <span class="s">{ICLR'25 -- CVPR'25 -- ICCV'25 -- NeurIPS'25}</span>
<span class="p">}</span></code></pre></figure>
</div>
<!-- Hidden bibtex block -->
<div class="submissions hidden">
<p>ICLR'25 -- CVPR'25 -- ICCV'25 -- NeurIPS'25</p>
</div>
</div>
</div>
</li>
<li>
<div class="row">
<!-- <div class="col-sm-1 abbr">
<div>
<abbr class="badge">ICML</abbr>
</div>
</div> -->
<div id="shojaee2025llmsrbench" class="col-sm-12">
<div class="title">
<span class="abbr">
<abbr class="badge">ICML</abbr>
</span>
<span class="d-inline-block" tabindex="0" data-toggle="tooltip" title="Majority of work done or significant contribution by MAIL/SAIL members!">
<abbr class="badge bg-danger"><i class="bi bi-check2"></i></abbr>
</span>
<abbr class="badge" style="background-color: #04a96d;">ORAL</abbr>
LLM-SRBench: A New Benchmark for Scientific Equation Discovery with Large Language Models
</div>
<div class="author">
Parshin Shojaee,
<span style="font-weight: bold;">Ngoc-Hieu Nguyen</span>,
Kazem Meidani,
Amir Barati Farimani,
<u>Khoa D Doan</u>,
and <a href="https://people.cs.vt.edu/reddy/index.html" target="_blank">Chandan K Reddy</a>
</div>
<div class="periodical">
<em>In International Conference on Machine Learning</em>
2025
</div>
<div class="links">
<a class="abstract btn btn-sm z-depth-0" role="button">Abs</a>
<a class="bibtex btn btn-sm z-depth-0" role="button">Bib</a>
<a href="https://arxiv.org/abs/2504.10415" class="btn btn-sm z-depth-0" role="button" target="_blank">PDF</a>
<a href="https://github.com/deep-symbolic-mathematics/llm-srbench" class="btn btn-sm z-depth-0" role="button" target="_blank">Code</a>
<a href="https://huggingface.co/datasets/nnheui/llm-srbench" class="btn btn-sm z-depth-0" role="button" target="_blank">Data</a>
<a class="submissions btn btn-sm z-depth-0" role="button">Submission History</a>
</div>
<!-- Hidden abstract block -->
<div class="abstract hidden">
<p>Scientific equation discovery is a fundamental task in the history of scientific progress, enabling the derivation of laws governing natural phenomena. Recently, Large Language Models (LLMs) have gained interest for this task due to their potential to leverage embedded scientific knowledge for hypothesis generation. However, evaluating the true discovery capabilities of these methods remains challenging, as existing benchmarks often rely on common equations that are susceptible to memorization by LLMs, leading to inflated performance metrics that do not reflect discovery. In this paper, we introduce LLM-SRBench, a comprehensive benchmark with 239 challenging problems across four scientific domains specifically designed to evaluate LLM-based scientific equation discovery methods while preventing trivial memorization. Our benchmark comprises two main categories: LSR-Transform, which transforms common physical models into less common mathematical representations to test reasoning beyond memorized forms, and LSR-Synth, which introduces synthetic, discovery-driven problems requiring data-driven reasoning. Through extensive evaluation of several state-of-the-art methods, using both open and closed LLMs, we find that the best-performing system so far achieves only 31.5% symbolic accuracy. These findings highlight the challenges of scientific equation discovery, positioning LLM-SRBench as a valuable resource for future research.</p>
</div>
<!-- Hidden bibtex block -->
<div class="bibtex hidden">
<figure class="highlight"><pre><code class="language-bibtex" data-lang="bibtex"><span class="nc">@inproceedings</span><span class="p">{</span><span class="nl">shojaee2025llmsrbench</span><span class="p">,</span>
<span class="na">title</span> <span class="p">=</span> <span class="s">{LLM-SRBench: A New Benchmark for Scientific Equation Discovery with Large Language Models}</span><span class="p">,</span>
<span class="na">author</span> <span class="p">=</span> <span class="s">{Shojaee, Parshin and Nguyen, Ngoc-Hieu and Meidani, Kazem and Farimani, Amir Barati and Doan, Khoa D and Reddy, Chandan K}</span><span class="p">,</span>
<span class="na">booktitle</span> <span class="p">=</span> <span class="s">{International Conference on Machine Learning}</span><span class="p">,</span>
<span class="na">year</span> <span class="p">=</span> <span class="s">{2025}</span><span class="p">,</span>
<span class="na">note</span> <span class="p">=</span> <span class="s">{ORAL}</span><span class="p">,</span>
<span class="na">bibtex_show</span> <span class="p">=</span> <span class="s">{true}</span><span class="p">,</span>
<span class="na">abbr</span> <span class="p">=</span> <span class="s">{ICML}</span><span class="p">,</span>
<span class="na">pdf</span> <span class="p">=</span> <span class="s">{https://arxiv.org/abs/2504.10415}</span><span class="p">,</span>
<span class="na">selected</span> <span class="p">=</span> <span class="s">{true}</span><span class="p">,</span>
<span class="na">code</span> <span class="p">=</span> <span class="s">{https://github.com/deep-symbolic-mathematics/llm-srbench}</span><span class="p">,</span>
<span class="na">data</span> <span class="p">=</span> <span class="s">{https://huggingface.co/datasets/nnheui/llm-srbench}</span><span class="p">,</span>
<span class="na">submissions</span> <span class="p">=</span> <span class="s">{ICML'25}</span>
<span class="p">}</span></code></pre></figure>
</div>
<!-- Hidden bibtex block -->
<div class="submissions hidden">
<p>ICML'25</p>
</div>
</div>
</div>
</li>
<li>
<div class="row">
<!-- <div class="col-sm-1 abbr">
<div>
<abbr class="badge">ICLR</abbr>
</div>
</div> -->
<div id="nguyen2024wicked" class="col-sm-12">
<div class="title">
<span class="abbr">
<abbr class="badge">ICLR</abbr>
</span>
<span class="d-inline-block" tabindex="0" data-toggle="tooltip" title="Majority of work done or significant contribution by MAIL/SAIL members!">
<abbr class="badge bg-danger"><i class="bi bi-check2"></i></abbr>
</span>
Wicked Oddities: Selectively Poisoning for Effective Clean-Label Backdoor Attacks
</div>
<div class="author">
<span style="font-weight: bold;">Quang H Nguyen</span>,
<span style="font-weight: bold;">Ngoc-Hieu Nguyen</span>,
<a href="https://the-anhta.github.io/" target="_blank">The-Anh Ta</a>,
<a href="https://thanhnguyentang.github.io/" target="_blank">Thanh Nguyen-Tang</a>,
<a href="https://sail-research.com/" target="_blank">Kok-Seng Wong</a>,
<a href="https://scholar.google.com/citations?user=xZU08d0AAAAJ&hl=en" target="_blank">Hoang Thanh-Tung</a>,
and <u>Khoa D Doan</u>
</div>
<div class="periodical">
<em>In The Twelfth International Conference on Learning Representations</em>
2025
</div>
<div class="links">
<a class="abstract btn btn-sm z-depth-0" role="button">Abs</a>
<a class="bibtex btn btn-sm z-depth-0" role="button">Bib</a>
<a href="https://openreview.net/forum?id=1Z3C49JQVf" class="btn btn-sm z-depth-0" role="button" target="_blank">PDF</a>
<a href="https://github.com/mail-research/wicked-oddities-backdoor" class="btn btn-sm z-depth-0" role="button" target="_blank">Code</a>
<a class="submissions btn btn-sm z-depth-0" role="button">Submission History</a>
</div>
<!-- Hidden abstract block -->
<div class="abstract hidden">
<p>Deep neural networks are vulnerable to backdoor attacks, a type of adversarial attack that poisons the training data to manipulate the behavior of models trained on such data. Clean-label attacks are a more stealthy form of backdoor attacks that can perform the attack without changing the labels of poisoned data. Early works on clean-label attacks added triggers to a random subset of the training set, ignoring the fact that samples contribute unequally to the attack's success. This results in high poisoning rates and low attack success rates. To alleviate the problem, several supervised learning-based sample selection strategies have been proposed. However, these methods assume access to the entire labeled training set and require training, which is expensive and may not always be practical. This work studies a new and more practical (but also more challenging) threat model where the attacker only provides data for the target class (e.g., in face recognition systems) and has no knowledge of the victim model or any other classes in the training set. We study different strategies for selectively poisoning a small set of training samples in the target class to boost the attack success rate in this setting. Our threat model poses a serious threat in training machine learning models with third-party datasets, since the attack can be performed effectively with limited information. Experiments on benchmark datasets illustrate the effectiveness of our strategies in improving clean-label backdoor attacks.</p>
</div>
<!-- Hidden bibtex block -->
<div class="bibtex hidden">
<figure class="highlight"><pre><code class="language-bibtex" data-lang="bibtex"><span class="nc">@inproceedings</span><span class="p">{</span><span class="nl">nguyen2024wicked</span><span class="p">,</span>
<span class="na">title</span> <span class="p">=</span> <span class="s">{Wicked Oddities: Selectively Poisoning for Effective Clean-Label Backdoor Attacks}</span><span class="p">,</span>
<span class="na">author</span> <span class="p">=</span> <span class="s">{Nguyen, Quang H and Nguyen, Ngoc-Hieu and Ta, The-Anh and Nguyen-Tang, Thanh and Wong, Kok-Seng and Thanh-Tung, Hoang and Doan, Khoa D}</span><span class="p">,</span>
<span class="na">booktitle</span> <span class="p">=</span> <span class="s">{The Twelfth International Conference on Learning Representations}</span><span class="p">,</span>
<span class="na">year</span> <span class="p">=</span> <span class="s">{2025}</span><span class="p">,</span>
<span class="na">bibtex_show</span> <span class="p">=</span> <span class="s">{true}</span><span class="p">,</span>
<span class="na">abbr</span> <span class="p">=</span> <span class="s">{ICLR}</span><span class="p">,</span>
<span class="na">pdf</span> <span class="p">=</span> <span class="s">{https://openreview.net/forum?id=1Z3C49JQVf}</span><span class="p">,</span>
<span class="na">selected</span> <span class="p">=</span> <span class="s">{true}</span><span class="p">,</span>
<span class="na">code</span> <span class="p">=</span> <span class="s">{https://github.com/mail-research/wicked-oddities-backdoor}</span><span class="p">,</span>
<span class="na">submissions</span> <span class="p">=</span> <span class="s">{NeurIPS-W'23 -- ICML'24 -- NeurIPS'24 -- ICLR'25}</span>
<span class="p">}</span></code></pre></figure>
</div>
<!-- Hidden bibtex block -->
<div class="submissions hidden">
<p>NeurIPS-W'23 -- ICML'24 -- NeurIPS'24 -- ICLR'25</p>
</div>
</div>
</div>
</li>
<li>
<div class="row">
<!-- <div class="col-sm-1 abbr">
<div>
<abbr class="badge">ECCV</abbr>
</div>
</div> -->
<div id="pham2024SBL" class="col-sm-12">
<div class="title">
<span class="abbr">
<abbr class="badge">ECCV</abbr>
</span>
<span class="d-inline-block" tabindex="0" data-toggle="tooltip" title="Majority of work done or significant contribution by MAIL/SAIL members!">
<abbr class="badge bg-danger"><i class="bi bi-check2"></i></abbr>
</span>
<abbr class="badge" style="background-color: #04a96d;">ORAL</abbr>
Flatness-aware Sequential Learning Generates Resilient Backdoors
</div>
<div class="author">
<span style="font-weight: bold;">Hoang V Pham</span>,
<a href="https://the-anhta.github.io/" target="_blank">The-Anh Ta</a>,
Anh Tran,
and <u>Khoa D Doan</u>
</div>
<div class="periodical">
<em>In European Conference on Computer Vision</em>
2024
</div>
<div class="links">
<a class="abstract btn btn-sm z-depth-0" role="button">Abs</a>
<a class="bibtex btn btn-sm z-depth-0" role="button">Bib</a>
<a href="https://www.arxiv.org/abs/2407.14738" class="btn btn-sm z-depth-0" role="button" target="_blank">PDF</a>
<a href="https://github.com/mail-research/SBL-resilient-backdoors" class="btn btn-sm z-depth-0" role="button" target="_blank">Code</a>
<a href="https://docs.google.com/presentation/d/1YEyQDSBardXdHCv-qBmxZnBwnXMAtRKZfWsawknowFs/edit?usp=sharing" class="btn btn-sm z-depth-0" role="button" target="_blank">Slides</a>
<a class="submissions btn btn-sm z-depth-0" role="button">Submission History</a>
</div>
<!-- Hidden abstract block -->
<div class="abstract hidden">
<p>Recently, backdoor attacks have become an emerging threat to the security of machine learning models. From the adversary's perspective, the implanted backdoors should be resistant to defensive algorithms, but some recently proposed fine-tuning defenses can remove these backdoors with notable efficacy. This is mainly due to the catastrophic forgetting (CF) property of deep neural networks. This paper counters CF of backdoors by leveraging continual learning (CL) techniques. We begin by investigating the connectivity between a backdoored and fine-tuned model in the loss landscape. Our analysis confirms that fine-tuning defenses, especially the more advanced ones, can easily push a poisoned model out of the backdoor regions, making it forget all about the backdoors. Based on this finding, we re-formulate backdoor training through the lens of CL and propose a novel framework, named Sequential Backdoor Learning (SBL), that can generate resilient backdoors. This framework separates the backdoor poisoning process into two tasks: the first task learns a backdoored model, while the second task, based on the CL principles, moves it to a backdoored region resistant to fine-tuning. We additionally propose to seek flatter backdoor regions via a sharpness-aware minimizer in the framework, further strengthening the durability of the implanted backdoor. Finally, we demonstrate the effectiveness of our method through extensive empirical experiments on several benchmark datasets in the backdoor domain. The source code is available at https://github.com/mail-research/SBL-resilient-backdoors</p>
</div>
<!-- Hidden bibtex block -->
<div class="bibtex hidden">
<figure class="highlight"><pre><code class="language-bibtex" data-lang="bibtex"><span class="nc">@inproceedings</span><span class="p">{</span><span class="nl">pham2024SBL</span><span class="p">,</span>
<span class="na">title</span> <span class="p">=</span> <span class="s">{Flatness-aware Sequential Learning Generates Resilient Backdoors}</span><span class="p">,</span>
<span class="na">author</span> <span class="p">=</span> <span class="s">{Pham, Hoang V and Ta, The-Anh and Tran, Anh and Doan, Khoa D}</span><span class="p">,</span>
<span class="na">booktitle</span> <span class="p">=</span> <span class="s">{European Conference on Computer Vision}</span><span class="p">,</span>
<span class="na">year</span> <span class="p">=</span> <span class="s">{2024}</span><span class="p">,</span>
<span class="na">note</span> <span class="p">=</span> <span class="s">{ORAL}</span><span class="p">,</span>
<span class="na">code</span> <span class="p">=</span> <span class="s">{https://github.com/mail-research/SBL-resilient-backdoors}</span><span class="p">,</span>
<span class="na">bibtex_show</span> <span class="p">=</span> <span class="s">{true}</span><span class="p">,</span>
<span class="na">abbr</span> <span class="p">=</span> <span class="s">{ECCV}</span><span class="p">,</span>
<span class="na">pdf</span> <span class="p">=</span> <span class="s">{https://www.arxiv.org/abs/2407.14738}</span><span class="p">,</span>
<span class="na">slides</span> <span class="p">=</span> <span class="s">{https://docs.google.com/presentation/d/1YEyQDSBardXdHCv-qBmxZnBwnXMAtRKZfWsawknowFs/edit?usp=sharing}</span><span class="p">,</span>
<span class="na">submissions</span> <span class="p">=</span> <span class="s">{CVPR'24 -- ECCV'24}</span><span class="p">,</span>
<span class="na">selected</span> <span class="p">=</span> <span class="s">{true}</span>
<span class="p">}</span></code></pre></figure>
</div>
<!-- Hidden bibtex block -->
<div class="submissions hidden">
<p>CVPR'24 -- ECCV'24</p>
</div>
</div>
</div>
</li>
<li>
<div class="row">
<!-- <div class="col-sm-1 abbr">
<div>
<abbr class="badge">ECCV</abbr>
</div>
</div> -->
<div id="huynh2024quantizedbd" class="col-sm-12">
<div class="title">
<span class="abbr">
<abbr class="badge">ECCV</abbr>
</span>
Data Poisoning Quantization Backdoor Attack
</div>
<div class="author">
Tran Huynh,
Anh Tran,
<u>Khoa D Doan</u>,
and <a href="https://scholar.google.com.au/citations?user=KcUuEKsAAAAJ&hl=en" target="_blank">Tung Pham</a>
</div>
<div class="periodical">
<em>In European Conference on Computer Vision</em>
2024
</div>
<div class="links">
<a class="abstract btn btn-sm z-depth-0" role="button">Abs</a>
<a class="bibtex btn btn-sm z-depth-0" role="button">Bib</a>
<a href="https://www.ecva.net/papers/eccv_2024/papers_ECCV/papers/11142.pdf" class="btn btn-sm z-depth-0" role="button" target="_blank">PDF</a>
<a class="submissions btn btn-sm z-depth-0" role="button">Submission History</a>
</div>
<!-- Hidden abstract block -->
<div class="abstract hidden">
<p>Deep learning (DL) models are often large and require a lot of computing power. Hence, model quantization is frequently used to reduce their size and complexity, making them more suitable for deployment on edge devices or achieving real-time performance. It has been previously shown that standard quantization frameworks can be exploited to activate the backdoor in a DL model. This means that an attacker could create a hijacked model that appears normal and free from backdoors (even when examined by state-of-the-art defenses), but when it is quantized, the backdoor is activated, and the attacker can control the model’s output. Existing backdoor attack methods on quantization models require full access to the victim model, which might not hold in practice. In this work, we focus on designing a novel quantization backdoor based on data poisoning, which requires zero knowledge of the target model. The key component is a trigger pattern generator, which is trained together with a surrogate model in an alternating manner. The attack’s effectiveness is tested on multiple benchmark datasets, including CIFAR10, CelebA, and ImageNet10, as well as state-of-the-art backdoor defenses.</p>
</div>
<!-- Hidden bibtex block -->
<div class="bibtex hidden">
<figure class="highlight"><pre><code class="language-bibtex" data-lang="bibtex"><span class="nc">@inproceedings</span><span class="p">{</span><span class="nl">huynh2024quantizedbd</span><span class="p">,</span>
<span class="na">title</span> <span class="p">=</span> <span class="s">{Data Poisoning Quantization Backdoor Attack}</span><span class="p">,</span>
<span class="na">author</span> <span class="p">=</span> <span class="s">{Huynh, Tran and Tran, Anh and Doan, Khoa D and Pham, Tung}</span><span class="p">,</span>
<span class="na">booktitle</span> <span class="p">=</span> <span class="s">{European Conference on Computer Vision}</span><span class="p">,</span>
<span class="na">year</span> <span class="p">=</span> <span class="s">{2024}</span><span class="p">,</span>
<span class="na">bibtex_show</span> <span class="p">=</span> <span class="s">{true}</span><span class="p">,</span>
<span class="na">abbr</span> <span class="p">=</span> <span class="s">{ECCV}</span><span class="p">,</span>
<span class="na">pdf</span> <span class="p">=</span> <span class="s">{https://www.ecva.net/papers/eccv_2024/papers_ECCV/papers/11142.pdf}</span><span class="p">,</span>
<span class="na">submissions</span> <span class="p">=</span> <span class="s">{CVPR'24 -- ECCV'24}</span><span class="p">,</span>
<span class="na">selected</span> <span class="p">=</span> <span class="s">{true}</span>
<span class="p">}</span></code></pre></figure>
</div>
<!-- Hidden bibtex block -->
<div class="submissions hidden">
<p>CVPR'24 -- ECCV'24</p>
</div>
</div>
</div>
</li>
<li>
<div class="row">
<!-- <div class="col-sm-1 abbr">
<div>
<abbr class="badge">PREPRINT</abbr>
</div>
</div> -->
<div id="nguyen2024metallm" class="col-sm-12">
<div class="title">
<span class="abbr">
<abbr class="badge" style="background-color: black;">PREPRINT</abbr>
</span>
<span class="d-inline-block" tabindex="0" data-toggle="tooltip" title="Majority of work done or significant contribution by MAIL/SAIL members!">
<abbr class="badge bg-danger"><i class="bi bi-check2"></i></abbr>
</span>
MetaLLM: A High-performant and Cost-efficient Dynamic Framework for Wrapping LLMs
</div>
<div class="author">
<span style="font-weight: bold;">Quang H Nguyen</span>,
<span style="font-weight: bold;">Cao-Duy Hoang</span>,
Juliette Decugis,
<a href="https://gurdaspuriya.github.io/" target="_blank">Saurav Manchanda</a>,
<a href="https://niteshchawla.nd.edu/" target="_blank">Nitesh V Chawla</a>,
and <u>Khoa D Doan</u>
</div>
<div class="periodical">
<em></em>
2024
</div>
<div class="links">
<a class="abstract btn btn-sm z-depth-0" role="button">Abs</a>
<a class="bibtex btn btn-sm z-depth-0" role="button">Bib</a>
<a href="https://arxiv.org/abs/2407.10834" class="btn btn-sm z-depth-0" role="button" target="_blank">PDF</a>
<a href="https://github.com/mail-research/MetaLLM-wrapper/" class="btn btn-sm z-depth-0" role="button" target="_blank">Code</a>
</div>
<!-- Hidden abstract block -->
<div class="abstract hidden">
<p>The rapid progress in machine learning (ML) has brought forth many large language models (LLMs) that excel in various tasks and areas. These LLMs come with different abilities and costs in terms of computation or pricing. Since the demand for each query can vary, e.g., because of the queried domain or its complexity, defaulting to one LLM in an application is not usually the best choice, whether it is the biggest, priciest, or even the one with the best average test performance. Consequently, picking the right LLM that is both accurate and cost-effective for an application remains a challenge. In this paper, we introduce MetaLLM, a framework that dynamically and intelligently routes each query to the optimal LLM (among several available LLMs) for classification tasks, achieving significantly improved accuracy and cost-effectiveness. By framing the selection problem as a multi-armed bandit, MetaLLM balances prediction accuracy and cost efficiency under uncertainty. Our experiments, conducted on popular LLM platforms such as OpenAI's GPT models, Amazon's Titan, Anthropic's Claude, and Meta's LLaMa, showcase MetaLLM's efficacy in real-world scenarios, laying the groundwork for future extensions beyond classification tasks.</p>
</div>
<!-- Hidden bibtex block -->
<div class="bibtex hidden">
<figure class="highlight"><pre><code class="language-bibtex" data-lang="bibtex"><span class="nc">@article</span><span class="p">{</span><span class="nl">nguyen2024metallm</span><span class="p">,</span>
<span class="na">title</span> <span class="p">=</span> <span class="s">{MetaLLM: A High-performant and Cost-efficient Dynamic Framework for Wrapping LLMs}</span><span class="p">,</span>
<span class="na">author</span> <span class="p">=</span> <span class="s">{Nguyen, Quang H and Hoang, Cao-Duy and Decugis, Juliette and Manchanda, Saurav and Chawla, Nitesh V and Doan, Khoa D}</span><span class="p">,</span>
<span class="na">year</span> <span class="p">=</span> <span class="s">{2024}</span><span class="p">,</span>
<span class="na">bibtex_show</span> <span class="p">=</span> <span class="s">{true}</span><span class="p">,</span>
<span class="na">abbr</span> <span class="p">=</span> <span class="s">{PREPRINT}</span><span class="p">,</span>
<span class="na">pdf</span> <span class="p">=</span> <span class="s">{https://arxiv.org/abs/2407.10834}</span><span class="p">,</span>
<span class="na">code</span> <span class="p">=</span> <span class="s">{https://github.com/mail-research/MetaLLM-wrapper/}</span><span class="p">,</span>
<span class="na">selected</span> <span class="p">=</span> <span class="s">{true}</span>
<span class="p">}</span></code></pre></figure>
</div>
<!-- Hidden bibtex block -->
</div>
</div>
</li>
<li>
<div class="row">
<!-- <div class="col-sm-1 abbr">
<div>
<abbr class="badge">UAI</abbr>
</div>
</div> -->
<div id="nguyen2024cold" class="col-sm-12">
<div class="title">
<span class="abbr">
<abbr class="badge">UAI</abbr>
</span>
Cold-start Recommendation by Personalized Embedding Region Elicitation
</div>
<div class="author">
<a href="https://hieunt91.github.io/" target="_blank">Hieu Trung Nguyen</a>,
<a href="https://duykhuongnguyen.github.io/" target="_blank">Duy Nguyen</a>,
<u>Khoa D Doan</u>,
and <a href="https://vietanhnguyen.net/" target="_blank">Viet Anh Nguyen</a>
</div>
<div class="periodical">
<em>In The Conference on Uncertainty in Artificial Intelligence</em>
2023
</div>
<div class="links">
<a class="abstract btn btn-sm z-depth-0" role="button">Abs</a>
<a class="bibtex btn btn-sm z-depth-0" role="button">Bib</a>
<a href="https://arxiv.org/pdf/2406.00973" class="btn btn-sm z-depth-0" role="button" target="_blank">PDF</a>
<a class="submissions btn btn-sm z-depth-0" role="button">Submission History</a>
</div>
<!-- Hidden abstract block -->
<div class="abstract hidden">
<p>Rating elicitation is a success element for recommender systems to perform well at cold-starting, in which the systems need to recommend items to a newly arrived user with no prior knowledge about the user's preference. Existing elicitation methods employ a fixed set of items to learn the user's preference and then infer the users' preferences on the remaining items. Using a fixed seed set can limit the performance of the recommendation system since the seed set is unlikely optimal for all new users with potentially diverse preferences. This paper addresses this challenge using a 2-phase, personalized elicitation scheme. First, the elicitation scheme asks users to rate a small set of popular items in a ``burn-in'' phase. Second, it sequentially asks the user to rate adaptive items to refine the preference and the user's representation. Throughout the process, the system represents the user's embedding value not by a point estimate but by a region estimate. The value of information obtained by asking the user's rating on an item is quantified by the distance from the region center embedding space that contains with high confidence the true embedding value of the user. Finally, the recommendations are successively generated by considering the preference region of the user. We show that each subproblem in the elicitation scheme can be efficiently implemented. Further, we empirically demonstrate the effectiveness of the proposed method against existing rating-elicitation methods on several prominent datasets.</p>
</div>
<!-- Hidden bibtex block -->
<div class="bibtex hidden">
<figure class="highlight"><pre><code class="language-bibtex" data-lang="bibtex"><span class="nc">@inproceedings</span><span class="p">{</span><span class="nl">nguyen2024cold</span><span class="p">,</span>
<span class="na">title</span> <span class="p">=</span> <span class="s">{Cold-start Recommendation by Personalized Embedding Region Elicitation}</span><span class="p">,</span>
<span class="na">author</span> <span class="p">=</span> <span class="s">{Nguyen, Hieu Trung and Nguyen, Duy and Doan, Khoa D and Nguyen, Viet Anh}</span><span class="p">,</span>
<span class="na">booktitle</span> <span class="p">=</span> <span class="s">{The Conference on Uncertainty in Artificial Intelligence}</span><span class="p">,</span>
<span class="na">year</span> <span class="p">=</span> <span class="s">{2023}</span><span class="p">,</span>
<span class="na">bibtex_show</span> <span class="p">=</span> <span class="s">{true}</span><span class="p">,</span>
<span class="na">abbr</span> <span class="p">=</span> <span class="s">{UAI}</span><span class="p">,</span>
<span class="na">pdf</span> <span class="p">=</span> <span class="s">{https://arxiv.org/pdf/2406.00973}</span><span class="p">,</span>
<span class="na">submissions</span> <span class="p">=</span> <span class="s">{RecSys'23 -- CIKM'23 -- AAAI'24 -- UAI'24}</span><span class="p">,</span>
<span class="na">selected</span> <span class="p">=</span> <span class="s">{true}</span>
<span class="p">}</span></code></pre></figure>
</div>
<!-- Hidden bibtex block -->
<div class="submissions hidden">
<p>RecSys'23 -- CIKM'23 -- AAAI'24 -- UAI'24</p>
</div>
</div>
</div>
</li>
<li>
<div class="row">
<!-- <div class="col-sm-1 abbr">
<div>
<abbr class="badge">PREPRINT</abbr>
</div>
</div> -->
<div id="yang2023synthesizing" class="col-sm-12">
<div class="title">
<span class="abbr">
<abbr class="badge" style="background-color: black;">PREPRINT</abbr>
</span>
<span class="d-inline-block" tabindex="0" data-toggle="tooltip" title="Majority of work done or significant contribution by MAIL/SAIL members!">
<abbr class="badge bg-danger"><i class="bi bi-check2"></i></abbr>
</span>
Synthesizing Physical Backdoor Datasets: An Automated Framework Leveraging Deep Generative Models
</div>
<div class="author">
<span style="font-weight: bold;">Sze Jue Yang</span>,
<span style="font-weight: bold;">Chinh D La</span>,
<span style="font-weight: bold;">Quang H Nguyen</span>,
<a href="https://www.cs.cornell.edu/~eugene/" target="_blank">Eugene Bagdasaryan</a>,
<a href="https://sail-research.com/" target="_blank">Kok-Seng Wong</a>,
<a href="https://scholar.google.com/citations?user=FYZ5ODQAAAAJ&hl=en" target="_blank">Anh T Tran</a>,
<a href="http://cs-chan.com/" target="_blank">Chee Seng Chan</a>,
and <u>Khoa D Doan</u>
</div>
<div class="periodical">
<em></em>
2024
</div>
<div class="links">
<a class="abstract btn btn-sm z-depth-0" role="button">Abs</a>
<a class="bibtex btn btn-sm z-depth-0" role="button">Bib</a>
<a href="https://arxiv.org/abs/2312.03419" class="btn btn-sm z-depth-0" role="button" target="_blank">PDF</a>
<a href="https://github.com/mail-research/synthetic-physical-backdoor-datasets" class="btn btn-sm z-depth-0" role="button" target="_blank">Code</a>
</div>
<!-- Hidden abstract block -->
<div class="abstract hidden">
<p>Backdoor attacks, representing an emerging threat to the integrity of deep neural networks, have garnered significant attention due to their ability to compromise deep learning systems clandestinely. While numerous backdoor attacks occur within the digital realm, their practical implementation in real-world prediction systems remains limited and vulnerable to disturbances in the physical world. Consequently, this limitation has given rise to the development of physical backdoor attacks, where trigger objects manifest as physical entities within the real world. However, creating the requisite dataset to train or evaluate a physical backdoor model is a daunting task, limiting the backdoor researchers and practitioners from studying such physical attack scenarios. This paper unleashes a recipe that empowers backdoor researchers to effortlessly create a malicious, physical backdoor dataset based on advances in generative modeling. Particularly, this recipe involves 3 automatic modules: suggesting the suitable physical triggers, generating the poisoned candidate samples (either by synthesizing new samples or editing existing clean samples), and finally refining for the most plausible ones. As such, it effectively mitigates the perceived complexity associated with creating a physical backdoor dataset, transforming it from a daunting task into an attainable objective. Extensive experiment results show that datasets created by our “recipe” enable adversaries to achieve an impressive attack success rate on real physical world data and exhibit similar properties compared to previous physical backdoor attack studies. This paper offers researchers a valuable toolkit for studies of physical backdoors, all within the confines of</p>
</div>
<!-- Hidden bibtex block -->
<div class="bibtex hidden">
<figure class="highlight"><pre><code class="language-bibtex" data-lang="bibtex"><span class="nc">@article</span><span class="p">{</span><span class="nl">yang2023synthesizing</span><span class="p">,</span>
<span class="na">title</span> <span class="p">=</span> <span class="s">{Synthesizing Physical Backdoor Datasets: An Automated Framework Leveraging Deep Generative Models}</span><span class="p">,</span>
<span class="na">author</span> <span class="p">=</span> <span class="s">{Yang, Sze Jue and La, Chinh D and Nguyen, Quang H and Bagdasaryan, Eugene and Wong, Kok-Seng and Tran, Anh T and Chan, Chee Seng and Doan, Khoa D}</span><span class="p">,</span>
<span class="na">year</span> <span class="p">=</span> <span class="s">{2024}</span><span class="p">,</span>
<span class="na">bibtex_show</span> <span class="p">=</span> <span class="s">{true}</span><span class="p">,</span>
<span class="na">selected</span> <span class="p">=</span> <span class="s">{true}</span><span class="p">,</span>
<span class="na">abbr</span> <span class="p">=</span> <span class="s">{PREPRINT}</span><span class="p">,</span>
<span class="na">code</span> <span class="p">=</span> <span class="s">{https://github.com/mail-research/synthetic-physical-backdoor-datasets}</span><span class="p">,</span>
<span class="na">pdf</span> <span class="p">=</span> <span class="s">{https://arxiv.org/abs/2312.03419}</span>
<span class="p">}</span></code></pre></figure>
</div>
<!-- Hidden bibtex block -->
</div>
</div>
</li>
<li>
<div class="row">
<!-- <div class="col-sm-1 abbr">
<div>
<abbr class="badge">ACL-Findings</abbr>
</div>
</div> -->
<div id="hoang2023advfooler" class="col-sm-12">
<div class="title">
<span class="abbr">
<abbr class="badge">ACL-Findings</abbr>
</span>
<span class="d-inline-block" tabindex="0" data-toggle="tooltip" title="Majority of work done or significant contribution by MAIL/SAIL members!">
<abbr class="badge bg-danger"><i class="bi bi-check2"></i></abbr>
</span>
Fooling the Textual Fooler via Randomizing Latent Representations
</div>
<div class="author">
<span style="font-weight: bold;">Cao-Duy Hoang</span>,
<span style="font-weight: bold;">Quang H Nguyen</span>,
<a href="https://gurdaspuriya.github.io/" target="_blank">Saurav Manchanda</a>,
<a href="https://v-mipeng.github.io/" target="_blank">Minlong Peng</a>,
<a href="https://sail-research.com/" target="_blank">Kok-Seng Wong</a>,
and <u>Khoa D Doan</u>
</div>
<div class="periodical">
<em>In Findings of the Association for Computational Linguistics</em>
2024
</div>
<div class="links">
<a class="abstract btn btn-sm z-depth-0" role="button">Abs</a>
<a class="bibtex btn btn-sm z-depth-0" role="button">Bib</a>
<a href="https://arxiv.org/abs/2310.01452" class="btn btn-sm z-depth-0" role="button" target="_blank">PDF</a>
<a href="https://github.com/mail-research/AdvFooler-text-defender" class="btn btn-sm z-depth-0" role="button" target="_blank">Code</a>
<a class="submissions btn btn-sm z-depth-0" role="button">Submission History</a>
</div>
<!-- Hidden abstract block -->
<div class="abstract hidden">
<p>Despite outstanding performance in a variety of NLP tasks, recent studies have revealed that NLP models are vulnerable to adversarial attacks that slightly perturb the input to cause the models to misbehave. Among these attacks, adversarial word-level perturbations are well-studied and effective attack strategies. Since these attacks work in black-box settings, they do not require access to the model architecture or model parameters and thus can be detrimental to existing NLP applications. To perform an attack, the adversary queries the victim model many times to determine the most important words in an input text and to replace these words with their corresponding synonyms. In this work, we propose a lightweight and attack-agnostic defense whose main goal is to perplex the process of generating an adversarial example in these query-based black-box attacks; that is to fool the textual fooler. This defense, named AdvFooler, works by randomizing the latent representation of the input at inference time. Different from existing defenses, AdvFooler does not necessitate additional computational overhead during training nor relies on assumptions about the potential adversarial perturbation set while having a negligible impact on the model's accuracy. Our theoretical and empirical analyses highlight the significance of robustness resulting from confusing the adversary via randomizing the latent space, as well as the impact of randomization on clean accuracy. Finally, we empirically demonstrate near state-of-the-art robustness of AdvFooler against representative adversarial word-level attacks on two benchmark datasets.</p>
</div>
<!-- Hidden bibtex block -->
<div class="bibtex hidden">
<figure class="highlight"><pre><code class="language-bibtex" data-lang="bibtex"><span class="nc">@inproceedings</span><span class="p">{</span><span class="nl">hoang2023advfooler</span><span class="p">,</span>
<span class="na">title</span> <span class="p">=</span> <span class="s">{Fooling the Textual Fooler via Randomizing Latent Representations}</span><span class="p">,</span>
<span class="na">author</span> <span class="p">=</span> <span class="s">{Hoang, Cao-Duy and Nguyen, Quang H and Manchanda, Saurav and Peng, Minlong and Wong, Kok-Seng and Doan, Khoa D}</span><span class="p">,</span>
<span class="na">year</span> <span class="p">=</span> <span class="s">{2024}</span><span class="p">,</span>
<span class="na">abbr</span> <span class="p">=</span> <span class="s">{ACL-Findings}</span><span class="p">,</span>
<span class="na">bibtex_show</span> <span class="p">=</span> <span class="s">{true}</span><span class="p">,</span>
<span class="na">booktitle</span> <span class="p">=</span> <span class="s">{Findings of the Association for Computational Linguistics}</span><span class="p">,</span>
<span class="na">pdf</span> <span class="p">=</span> <span class="s">{https://arxiv.org/abs/2310.01452}</span><span class="p">,</span>
<span class="na">code</span> <span class="p">=</span> <span class="s">{https://github.com/mail-research/AdvFooler-text-defender}</span><span class="p">,</span>
<span class="na">submissions</span> <span class="p">=</span> <span class="s">{EMNLP'24 -- ICLR'24 -- ACL'24}</span><span class="p">,</span>
<span class="na">selected</span> <span class="p">=</span> <span class="s">{true}</span>
<span class="p">}</span></code></pre></figure>
</div>
<!-- Hidden bibtex block -->
<div class="submissions hidden">
<p>EMNLP'24 -- ICLR'24 -- ACL'24</p>
</div>
</div>
</div>
</li>
<li>
<div class="row">
<!-- <div class="col-sm-1 abbr">
<div>
<abbr class="badge">ICLR</abbr>
</div>
</div> -->
<div id="nguyen2023randomized" class="col-sm-12">
<div class="title">
<span class="abbr">
<abbr class="badge">ICLR</abbr>
</span>
<span class="d-inline-block" tabindex="0" data-toggle="tooltip" title="Majority of work done or significant contribution by MAIL/SAIL members!">
<abbr class="badge bg-danger"><i class="bi bi-check2"></i></abbr>
</span>
Understanding the Robustness of Randomized Feature Defense Against Query-Based Adversarial Attacks
</div>
<div class="author">
<span style="font-weight: bold;">Quang H Nguyen</span>,
<a href="http://ylao.people.clemson.edu/" target="_blank">Yingjie Lao</a>,
<a href="https://scholar.google.com.au/citations?user=KcUuEKsAAAAJ&hl=en" target="_blank">Tung Pham</a>,
<a href="https://sail-research.com/" target="_blank">Kok-Seng Wong</a>,
and <u>Khoa D Doan</u>
</div>
<div class="periodical">
<em>In The Twelfth International Conference on Learning Representations</em>
2024
</div>
<div class="links">
<a class="abstract btn btn-sm z-depth-0" role="button">Abs</a>
<a class="bibtex btn btn-sm z-depth-0" role="button">Bib</a>
<a href="https://openreview.net/forum?id=vZ6r9GMT1n" class="btn btn-sm z-depth-0" role="button" target="_blank">PDF</a>
<a href="https://github.com/mail-research/randomized_defenses" class="btn btn-sm z-depth-0" role="button" target="_blank">Code</a>
<a class="submissions btn btn-sm z-depth-0" role="button">Submission History</a>
</div>
<!-- Hidden abstract block -->
<div class="abstract hidden">
<p>Recent works have shown that deep neural networks are vulnerable to adversarial examples that find samples close to the original image but can make the model misclassify. Even with access only to the model's output, an attacker can employ black-box attacks to generate such adversarial examples. In this work, we propose a simple and lightweight defense against black-box attacks by adding random noise to hidden features at intermediate layers of the model at inference time. Our theoretical analysis confirms that this method effectively enhances the model's resilience against both score-based and decision-based black-box attacks. Importantly, our defense does not necessitate adversarial training and has minimal impact on accuracy, rendering it applicable to any pre-trained model. Our analysis also reveals the significance of selectively adding noise to different parts of the model based on the gradient of the adversarial objective function, which can be varied during the attack. We demonstrate the robustness of our defense against multiple black-box attacks through extensive empirical experiments involving diverse models with various architectures.</p>
</div>
<!-- Hidden bibtex block -->
<div class="bibtex hidden">
<figure class="highlight"><pre><code class="language-bibtex" data-lang="bibtex"><span class="nc">@inproceedings</span><span class="p">{</span><span class="nl">nguyen2023randomized</span><span class="p">,</span>
<span class="na">title</span> <span class="p">=</span> <span class="s">{Understanding the Robustness of Randomized Feature Defense Against Query-Based Adversarial Attacks}</span><span class="p">,</span>
<span class="na">author</span> <span class="p">=</span> <span class="s">{Nguyen, Quang H and Lao, Yingjie and Pham, Tung and Wong, Kok-Seng and Doan, Khoa D}</span><span class="p">,</span>
<span class="na">booktitle</span> <span class="p">=</span> <span class="s">{The Twelfth International Conference on Learning Representations}</span><span class="p">,</span>
<span class="na">year</span> <span class="p">=</span> <span class="s">{2024}</span><span class="p">,</span>
<span class="na">bibtex_show</span> <span class="p">=</span> <span class="s">{true}</span><span class="p">,</span>
<span class="na">abbr</span> <span class="p">=</span> <span class="s">{ICLR}</span><span class="p">,</span>
<span class="na">pdf</span> <span class="p">=</span> <span class="s">{https://openreview.net/forum?id=vZ6r9GMT1n}</span><span class="p">,</span>
<span class="na">selected</span> <span class="p">=</span> <span class="s">{true}</span><span class="p">,</span>
<span class="na">code</span> <span class="p">=</span> <span class="s">{https://github.com/mail-research/randomized_defenses}</span><span class="p">,</span>
<span class="na">submissions</span> <span class="p">=</span> <span class="s">{NeurIPS'23 -- ICLR'24}</span>
<span class="p">}</span></code></pre></figure>
</div>
<!-- Hidden bibtex block -->
<div class="submissions hidden">
<p>NeurIPS'23 -- ICLR'24</p>
</div>
</div>
</div>
</li>
<li>
<div class="row">
<!-- <div class="col-sm-1 abbr">
<div>
<abbr class="badge">NeurIPS</abbr>
</div>
</div> -->
<div id="nguyen2024iba" class="col-sm-12">
<div class="title">
<span class="abbr">
<abbr class="badge">NeurIPS</abbr>
</span>
<span class="d-inline-block" tabindex="0" data-toggle="tooltip" title="Majority of work done or significant contribution by MAIL/SAIL members!">
<abbr class="badge bg-danger"><i class="bi bi-check2"></i></abbr>
</span>
Iba: Towards irreversible backdoor attacks in federated learning
</div>
<div class="author">
<span style="font-weight: bold;">Thuy Dung Nguyen</span>,
<span style="font-weight: bold;">Tuan M Nguyen</span>,
<a href="https://scholar.google.com/citations?user=FYZ5ODQAAAAJ&hl=en" target="_blank">Anh T Tran</a>,
<u>Khoa D Doan</u>,
and <a href="https://sail-research.com/" target="_blank">Kok-Seng Wong</a>
</div>
<div class="periodical">
<em>Advances in Neural Information Processing Systems</em>
2024
</div>
<div class="links">
<a class="abstract btn btn-sm z-depth-0" role="button">Abs</a>
<a class="bibtex btn btn-sm z-depth-0" role="button">Bib</a>
<a href="https://proceedings.neurips.cc/paper_files/paper/2023/hash/d0c6bc641a56bebee9d985b937307367-Abstract-Conference.html" class="btn btn-sm z-depth-0" role="button" target="_blank">PDF</a>
<a href="https://github.com/sail-research/iba" class="btn btn-sm z-depth-0" role="button" target="_blank">Code</a>
<a class="submissions btn btn-sm z-depth-0" role="button">Submission History</a>
</div>
<!-- Hidden abstract block -->
<div class="abstract hidden">
<p>Federated learning (FL) is a distributed learning approach that enables machine learning models to be trained on decentralized data without compromising end devices' personal, potentially sensitive data. However, the distributed nature and uninvestigated data intuitively introduce new security vulnerabilities, including backdoor attacks. In this scenario, an adversary implants backdoor functionality into the global model during training, which can be activated to cause the desired misbehaviors for any input with a specific adversarial pattern. Despite having remarkable success in triggering and distorting model behavior, prior backdoor attacks in FL often hold impractical assumptions, limited imperceptibility, and durability. Specifically, the adversary needs to control a sufficiently large fraction of clients or know the data distribution of other honest clients. In many cases, the trigger inserted is often visually apparent, and the backdoor effect is quickly diluted if the adversary is removed from the training process. To address these limitations, we propose a novel backdoor attack framework in FL, the Irreversible Backdoor Attack (IBA), that jointly learns the optimal and visually stealthy trigger and then gradually implants the backdoor into a global model. This approach allows the adversary to execute a backdoor attack that can evade both human and machine inspections. Additionally, we enhance the efficiency and durability of the proposed attack by selectively poisoning the model's parameters that are least likely updated by the main task's learning process and constraining the poisoned model update to the vicinity of the global model. Finally, we evaluate the proposed attack framework on several benchmark datasets, including MNIST, CIFAR-10, and Tiny ImageNet, and achieved high success rates while simultaneously bypassing existing backdoor defenses and achieving a more durable backdoor effect compared to other backdoor attacks. Overall, IBA offers a more effective, stealthy, and durable approach to backdoor attacks in FL.</p>
</div>
<!-- Hidden bibtex block -->
<div class="bibtex hidden">
<figure class="highlight"><pre><code class="language-bibtex" data-lang="bibtex"><span class="nc">@article</span><span class="p">{</span><span class="nl">nguyen2024iba</span><span class="p">,</span>
<span class="na">title</span> <span class="p">=</span> <span class="s">{Iba: Towards irreversible backdoor attacks in federated learning}</span><span class="p">,</span>
<span class="na">author</span> <span class="p">=</span> <span class="s">{Nguyen, Thuy Dung and Nguyen, Tuan M and Tran, Anh T and Doan, Khoa D and Wong, Kok-Seng}</span><span class="p">,</span>
<span class="na">journal</span> <span class="p">=</span> <span class="s">{Advances in Neural Information Processing Systems}</span><span class="p">,</span>
<span class="na">volume</span> <span class="p">=</span> <span class="s">{36}</span><span class="p">,</span>
<span class="na">year</span> <span class="p">=</span> <span class="s">{2024}</span><span class="p">,</span>
<span class="na">submissions</span> <span class="p">=</span> <span class="s">{NeurIPS'23}</span><span class="p">,</span>
<span class="na">bibtex_show</span> <span class="p">=</span> <span class="s">{true}</span><span class="p">,</span>
<span class="na">abbr</span> <span class="p">=</span> <span class="s">{NeurIPS}</span><span class="p">,</span>
<span class="na">pdf</span> <span class="p">=</span> <span class="s">{https://proceedings.neurips.cc/paper_files/paper/2023/hash/d0c6bc641a56bebee9d985b937307367-Abstract-Conference.html}</span><span class="p">,</span>
<span class="na">code</span> <span class="p">=</span> <span class="s">{https://github.com/sail-research/iba}</span><span class="p">,</span>
<span class="na">selected</span> <span class="p">=</span> <span class="s">{true}</span>
<span class="p">}</span></code></pre></figure>
</div>
<!-- Hidden bibtex block -->
<div class="submissions hidden">
<p>NeurIPS'23</p>
</div>
</div>
</div>
</li>
<li>
<div class="row">
<!-- <div class="col-sm-1 abbr">
<div>
<abbr class="badge">EAAI</abbr>
</div>
</div> -->
<div id="nguyen2024backdoor" class="col-sm-12">
<div class="title">
<span class="abbr">
<abbr class="badge">EAAI</abbr>
</span>
<span class="d-inline-block" tabindex="0" data-toggle="tooltip" title="Majority of work done or significant contribution by MAIL/SAIL members!">
<abbr class="badge bg-danger"><i class="bi bi-check2"></i></abbr>
</span>
Backdoor attacks and defenses in federated learning: Survey, challenges and future research directions
</div>
<div class="author">
<span style="font-weight: bold;">Thuy Dung Nguyen</span>,
<span style="font-weight: bold;">Tuan M Nguyen</span>,
Phi Le Nguyen,
Hieu H Pham,
<u>Khoa D Doan</u>,
and <a href="https://sail-research.com/" target="_blank">Kok-Seng Wong</a>
</div>
<div class="periodical">
<em>Engineering Applications of Artificial Intelligence</em>
2024
</div>
<div class="links">
<a class="abstract btn btn-sm z-depth-0" role="button">Abs</a>
<a class="bibtex btn btn-sm z-depth-0" role="button">Bib</a>
<a href="https://www.sciencedirect.com/science/article/pii/S0952197623013507" class="btn btn-sm z-depth-0" role="button" target="_blank">PDF</a>
<a class="submissions btn btn-sm z-depth-0" role="button">Submission History</a>
</div>
<!-- Hidden abstract block -->
<div class="abstract hidden">
<p>Federated learning (FL) is an approach within the realm of machine learning (ML) that allows the use of distributed data without compromising personal privacy. In FL, it becomes evident that the training data among participants frequently exhibit heterogeneous distribution characteristics. This inherent heterogeneity poses a substantial challenge for the orchestration server as it strives to assess the reliability of each local model update. Due to this challenge, FL becomes susceptible to various potential risks, with the ominous backdoor attack standing out as one of the most menacing threats. Backdoor attacks involve the insertion of malicious functionality into a targeted model through poisoned updates from malicious clients. These attacks can cause the global model to misbehave on specific inputs while appearing normal in other instances. Although the backdoor attacks received significant attention for their potential impact on practical deep learning applications, their exploration within the realm of FL remains limited. This survey seeks to address this gap by offering an all-encompassing examination of prevailing backdoor attack tactics and defenses in the context of FL. We include an exhaustive analysis of diverse approaches to provide a comprehensive understanding of this intricate landscape. Furthermore, we also discuss the challenges and potential future directions for attacks and defenses in the context of FL.</p>
</div>
<!-- Hidden bibtex block -->
<div class="bibtex hidden">
<figure class="highlight"><pre><code class="language-bibtex" data-lang="bibtex"><span class="nc">@article</span><span class="p">{</span><span class="nl">nguyen2024backdoor</span><span class="p">,</span>
<span class="na">title</span> <span class="p">=</span> <span class="s">{Backdoor attacks and defenses in federated learning: Survey, challenges and future research directions}</span><span class="p">,</span>
<span class="na">author</span> <span class="p">=</span> <span class="s">{Nguyen, Thuy Dung and Nguyen, Tuan M and Le Nguyen, Phi and Pham, Hieu H and Doan, Khoa D and Wong, Kok-Seng}</span><span class="p">,</span>
<span class="na">journal</span> <span class="p">=</span> <span class="s">{Engineering Applications of Artificial Intelligence}</span><span class="p">,</span>
<span class="na">volume</span> <span class="p">=</span> <span class="s">{127}</span><span class="p">,</span>
<span class="na">pages</span> <span class="p">=</span> <span class="s">{107166}</span><span class="p">,</span>
<span class="na">year</span> <span class="p">=</span> <span class="s">{2024}</span><span class="p">,</span>
<span class="na">publisher</span> <span class="p">=</span> <span class="s">{Elsevier}</span><span class="p">,</span>
<span class="na">submissions</span> <span class="p">=</span> <span class="s">{EAAI'24}</span><span class="p">,</span>
<span class="na">bibtex_show</span> <span class="p">=</span> <span class="s">{true}</span><span class="p">,</span>
<span class="na">abbr</span> <span class="p">=</span> <span class="s">{EAAI}</span><span class="p">,</span>
<span class="na">pdf</span> <span class="p">=</span> <span class="s">{https://www.sciencedirect.com/science/article/pii/S0952197623013507}</span><span class="p">,</span>
<span class="na">selected</span> <span class="p">=</span> <span class="s">{true}</span>
<span class="p">}</span></code></pre></figure>
</div>
<!-- Hidden bibtex block -->
<div class="submissions hidden">
<p>EAAI'24</p>
</div>
</div>
</div>
</li>
<li>
<div class="row">
<!-- <div class="col-sm-1 abbr">
<div>
<abbr class="badge">SIGIR</abbr>
</div>
</div> -->
<div id="doan2023flora" class="col-sm-12">
<div class="title">
<span class="abbr">
<abbr class="badge">SIGIR</abbr>
</span>
<span class="d-inline-block" tabindex="0" data-toggle="tooltip" title="Majority of work done or significant contribution by MAIL/SAIL members!">
<abbr class="badge bg-danger"><i class="bi bi-check2"></i></abbr>
</span>
Asymmetric Hashing for Fast Ranking via Neural Network Measures
</div>
<div class="author">
<u>Khoa D Doan</u>,
<a href="http://www.stat.ucla.edu/~jxie/" target="_blank">Shulong Tan</a>,
<a href="https://www.cs.rit.edu/~wjz/" target="_blank">Weijie Zhao</a>,
and <a href="http://research.baidu.com/People/index-view?id=111" target="_blank">Ping Li</a>
</div>
<div class="periodical">
<em>In 46th International ACM SIGIR Conference on Research and Development in Information Retrieval</em>
2023
</div>
<div class="links">
<a class="abstract btn btn-sm z-depth-0" role="button">Abs</a>
<a class="bibtex btn btn-sm z-depth-0" role="button">Bib</a>
<a href="https://dl.acm.org/doi/abs/10.1145/3539618.3591640" class="btn btn-sm z-depth-0" role="button" target="_blank">PDF</a>
<a class="submissions btn btn-sm z-depth-0" role="button">Submission History</a>
</div>
<!-- Hidden abstract block -->
<div class="abstract hidden">
<p>Fast item ranking is an important task in recommender systems. In previous works, graph-based Approximate Nearest Neighbor (ANN) approaches have demonstrated good performance on item ranking tasks with generic searching/matching measures (including complex measures such as neural network measures). However, since these ANN approaches must go through the neural measures several times during ranking, the computation is not practical if the neural measure is a large network. On the other hand, fast item ranking using existing hashing-based approaches, such as Locality Sensitive Hashing (LSH), only works with a limited set of measures, such as cosine and Euclidean distance, but not with general search measures such as neural networks. Given an arbitrary searching measure, previous learning-to-hash approaches are also not suitable to solve the fast item ranking problem since they can take a significant amount of time and computation to train the hash functions to approximate the searching measure due to a large number of possible training pairs in this problem. Hashing approaches, however, are attractive because they provide a principal and efficient way to retrieve candidate items. In this paper, we propose a simple and effective learning-to-hash approach for the fast item ranking problem that can be used to efficiently approximate any type of measure, including neural network measures. Specifically, we solve this problem with an asymmetric hashing framework based on discrete inner product fitting. We learn a pair of related hash functions that map heterogeneous objects (e.g., users and items) into a common discrete space where the inner product of their binary codes reveals their true similarity defined via the original searching measure. The fast ranking problem is reduced to an ANN search via this asymmetric hashing scheme. Then, we propose a sampling strategy to efficiently select relevant and contrastive samples to train the hashing model. We empirically validate the proposed method against the existing state-of-the-art fast item ranking methods in several combinations of non-linear searching functions and prominent datasets.</p>
</div>
<!-- Hidden bibtex block -->
<div class="bibtex hidden">
<figure class="highlight"><pre><code class="language-bibtex" data-lang="bibtex"><span class="nc">@inproceedings</span><span class="p">{</span><span class="nl">doan2023flora</span><span class="p">,</span>
<span class="na">bibtex_show</span> <span class="p">=</span> <span class="s">{true}</span><span class="p">,</span>
<span class="na">abbr</span> <span class="p">=</span> <span class="s">{SIGIR}</span><span class="p">,</span>
<span class="na">title</span> <span class="p">=</span> <span class="s">{Asymmetric Hashing for Fast Ranking via Neural Network Measures}</span><span class="p">,</span>
<span class="na">author</span> <span class="p">=</span> <span class="s">{Doan, Khoa D and Tan, Shulong and Zhao, Weijie and Li, Ping}</span><span class="p">,</span>
<span class="na">booktitle</span> <span class="p">=</span> <span class="s">{46th International ACM SIGIR Conference on Research and Development in Information Retrieval}</span><span class="p">,</span>
<span class="na">year</span> <span class="p">=</span> <span class="s">{2023}</span><span class="p">,</span>
<span class="na">submissions</span> <span class="p">=</span> <span class="s">{SIGIR'21 -- RecSys'21 -- WSDM'22 -- WWW'22 -- SIGIR'22 -- VLDB'23 -- CIKM'22 -- WWW'23 -- SIGIR'23}</span><span class="p">,</span>
<span class="na">pdf</span> <span class="p">=</span> <span class="s">{https://dl.acm.org/doi/abs/10.1145/3539618.3591640}</span><span class="p">,</span>
<span class="na">selected</span> <span class="p">=</span> <span class="s">{true}</span>
<span class="p">}</span></code></pre></figure>
</div>
<!-- Hidden bibtex block -->
<div class="submissions hidden">
<p>SIGIR'21 -- RecSys'21 -- WSDM'22 -- WWW'22 -- SIGIR'22 -- VLDB'23 -- CIKM'22 -- WWW'23 -- SIGIR'23</p>
</div>
</div>
</div>
</li>
<li>
<div class="row">
<!-- <div class="col-sm-1 abbr">
<div>
<abbr class="badge">AAAI</abbr>
</div>
</div> -->
<div id="doan2023bdvitt" class="col-sm-12">
<div class="title">
<span class="abbr">
<abbr class="badge">AAAI</abbr>
</span>
<span class="d-inline-block" tabindex="0" data-toggle="tooltip" title="Majority of work done or significant contribution by MAIL/SAIL members!">
<abbr class="badge bg-danger"><i class="bi bi-check2"></i></abbr>
</span>
Defending backdoor attacks on vision transformer via patch processing
</div>
<div class="author">
<u>Khoa D Doan</u>,
<a href="http://ylao.people.clemson.edu/" target="_blank">Yingjie Lao</a>,
and <a href="http://research.baidu.com/People/index-view?id=111" target="_blank">Ping Li</a>
</div>
<div class="periodical">
<em>In AAAI Conference on Artificial Intelligence</em>
2023
</div>
<div class="links">
<a class="abstract btn btn-sm z-depth-0" role="button">Abs</a>
<a class="bibtex btn btn-sm z-depth-0" role="button">Bib</a>
<a href="https://ojs.aaai.org/index.php/AAAI/article/view/25125" class="btn btn-sm z-depth-0" role="button" target="_blank">PDF</a>
<a class="submissions btn btn-sm z-depth-0" role="button">Submission History</a>
</div>
<!-- Hidden abstract block -->
<div class="abstract hidden">
<p>Vision Transformers (ViTs) have a radically different architecture with significantly less inductive bias than Convolutional Neural Networks. Along with the improvement in performance, security and robustness of ViTs are also of great importance to study. In contrast to many recent works that exploit the robustness of ViTs against adversarial examples, this paper investigates a representative causative attack, ie, backdoor. We first examine the vulnerability of ViTs against various backdoor attacks and find that ViTs are also quite vulnerable to existing attacks. However, we observe that the clean-data accuracy and backdoor attack success rate of ViTs respond distinctively to patch transformations before the positional encoding. Then, based on this finding, we propose an effective method for ViTs to defend both patch-based and blending-based trigger backdoor attacks via patch processing. The performances are evaluated on several benchmark datasets, including CIFAR10, GTSRB, and TinyImageNet, which show the proposedds defense is very successful in mitigating backdoor attacks for ViTs. To the best of our knowledge, this paper presents the first defensive strategy that utilizes a unique characteristic of ViTs against backdoor attacks.</p>
</div>
<!-- Hidden bibtex block -->
<div class="bibtex hidden">
<figure class="highlight"><pre><code class="language-bibtex" data-lang="bibtex"><span class="nc">@inproceedings</span><span class="p">{</span><span class="nl">doan2023bdvitt</span><span class="p">,</span>
<span class="na">abbr</span> <span class="p">=</span> <span class="s">{AAAI}</span><span class="p">,</span>
<span class="na">title</span> <span class="p">=</span> <span class="s">{Defending backdoor attacks on vision transformer via patch processing}</span><span class="p">,</span>
<span class="na">author</span> <span class="p">=</span> <span class="s">{Doan, Khoa D and Lao, Yingjie and Li, Ping}</span><span class="p">,</span>
<span class="na">booktitle</span> <span class="p">=</span> <span class="s">{AAAI Conference on Artificial Intelligence}</span><span class="p">,</span>
<span class="na">year</span> <span class="p">=</span> <span class="s">{2023}</span><span class="p">,</span>
<span class="na">submissions</span> <span class="p">=</span> <span class="s">{CVPR'22 -- ICCV'22 -- AAAI'23}</span><span class="p">,</span>
<span class="na">bibtex_show</span> <span class="p">=</span> <span class="s">{true}</span><span class="p">,</span>
<span class="na">selected</span> <span class="p">=</span> <span class="s">{true}</span><span class="p">,</span>
<span class="na">teaser</span> <span class="p">=</span> <span class="s">{doan2023bdvits.png}</span><span class="p">,</span>
<span class="na">pdf</span> <span class="p">=</span> <span class="s">{https://ojs.aaai.org/index.php/AAAI/article/view/25125}</span>
<span class="p">}</span></code></pre></figure>
</div>
<!-- Hidden bibtex block -->
<div class="submissions hidden">
<p>CVPR'22 -- ICCV'22 -- AAAI'23</p>
</div>
</div>
</div>
</li>
<li>
<div class="row">
<!-- <div class="col-sm-1 abbr">
<div>
<abbr class="badge">NeurIPS</abbr>
</div>
</div> -->
<div id="doan2022marksman" class="col-sm-12">
<div class="title">
<span class="abbr">
<abbr class="badge">NeurIPS</abbr>
</span>
<span class="d-inline-block" tabindex="0" data-toggle="tooltip" title="Majority of work done or significant contribution by MAIL/SAIL members!">
<abbr class="badge bg-danger"><i class="bi bi-check2"></i></abbr>
</span>
Marksman Backdoor: Backdoor Attacks with Arbitrary Target Class
</div>
<div class="author">
<u>Khoa D Doan</u>,
<a href="http://ylao.people.clemson.edu/" target="_blank">Yingjie Lao</a>,
and <a href="http://research.baidu.com/People/index-view?id=111" target="_blank">Ping Li</a>
</div>
<div class="periodical">
<em>In Thirty-Sixth Conference on Neural Information Processing Systems</em>
2022
</div>
<div class="links">
<a class="abstract btn btn-sm z-depth-0" role="button">Abs</a>
<a class="bibtex btn btn-sm z-depth-0" role="button">Bib</a>
<a href="https://openreview.net/forum?id=i-k6J4VkCDq" class="btn btn-sm z-depth-0" role="button" target="_blank">PDF</a>
<a href="https://github.com/khoadoan106/backdoor_attacks" class="btn btn-sm z-depth-0" role="button" target="_blank">Code</a>
<a href="https://nips.cc/media/neurips-2022/Slides/52924.pdf" class="btn btn-sm z-depth-0" role="button" target="_blank">Slides</a>
<a class="submissions btn btn-sm z-depth-0" role="button">Submission History</a>
</div>
<!-- Hidden abstract block -->
<div class="abstract hidden">
<p>In recent years, machine learning models have been shown to be vulnerable to backdoor attacks. Under such attacks, an adversary embeds a stealthy backdoor into the trained model such that the compromised models will behave normally on clean inputs but will misclassify according to the adversary's control on maliciously constructed input with a trigger. While these existing attacks are very effective, the adversary's capability is limited: given an input, these attacks can only cause the model to misclassify toward a single pre-defined or target class. In contrast, this paper exploits a novel backdoor attack with a much more powerful payload, denoted as Marksman, where the adversary can arbitrarily choose which target class the model will misclassify given any input during inference. To achieve this goal, we propose to represent the trigger function as a class-conditional generative model and to inject the backdoor in a constrained optimization framework, where the trigger function learns to generate an optimal trigger pattern to attack any target class at will while simultaneously embedding this generative backdoor into the trained model. Given the learned trigger-generation function, during inference, the adversary can specify an arbitrary backdoor attack target class, and an appropriate trigger causing the model to classify toward this target class is created accordingly. We show empirically that the proposed framework achieves high attack performance (e.g., 100% attack success rates in several experiments) while preserving the clean-data performance in several benchmark datasets, including MNIST, CIFAR10, GTSRB, and TinyImageNet. The proposed Marksman backdoor attack can also easily bypass existing backdoor defenses that were originally designed against backdoor attacks with a single target class. Our work takes another significant step toward understanding the extensive risks of backdoor attacks in practice.</p>
</div>
<!-- Hidden bibtex block -->
<div class="bibtex hidden">
<figure class="highlight"><pre><code class="language-bibtex" data-lang="bibtex"><span class="nc">@inproceedings</span><span class="p">{</span><span class="nl">doan2022marksman</span><span class="p">,</span>
<span class="na">bibtex_show</span> <span class="p">=</span> <span class="s">{true}</span><span class="p">,</span>
<span class="na">abbr</span> <span class="p">=</span> <span class="s">{NeurIPS}</span><span class="p">,</span>
<span class="na">title</span> <span class="p">=</span> <span class="s">{Marksman Backdoor: Backdoor Attacks with Arbitrary Target Class}</span><span class="p">,</span>
<span class="na">author</span> <span class="p">=</span> <span class="s">{Doan, Khoa D and Lao, Yingjie and Li, Ping}</span><span class="p">,</span>
<span class="na">booktitle</span> <span class="p">=</span> <span class="s">{Thirty-Sixth Conference on Neural Information Processing Systems}</span><span class="p">,</span>
<span class="na">year</span> <span class="p">=</span> <span class="s">{2022}</span><span class="p">,</span>
<span class="na">submissions</span> <span class="p">=</span> <span class="s">{NeurIPS'22}</span><span class="p">,</span>
<span class="na">teaser</span> <span class="p">=</span> <span class="s">{doan2022marksman.png}</span><span class="p">,</span>
<span class="na">pdf</span> <span class="p">=</span> <span class="s">{https://openreview.net/forum?id=i-k6J4VkCDq}</span><span class="p">,</span>
<span class="na">code</span> <span class="p">=</span> <span class="s">{https://github.com/khoadoan106/backdoor_attacks}</span><span class="p">,</span>
<span class="na">slides</span> <span class="p">=</span> <span class="s">{https://nips.cc/media/neurips-2022/Slides/52924.pdf}</span><span class="p">,</span>
<span class="na">selected</span> <span class="p">=</span> <span class="s">{true}</span>
<span class="p">}</span></code></pre></figure>
</div>
<!-- Hidden bibtex block -->
<div class="submissions hidden">
<p>NeurIPS'22</p>
</div>
</div>
</div>
</li>
<li>
<div class="row">
<!-- <div class="col-sm-1 abbr">
<div>
<abbr class="badge">CVPR</abbr>
</div>
</div> -->
<div id="doan2022hswd" class="col-sm-12">
<div class="title">
<span class="abbr">
<abbr class="badge">CVPR</abbr>
</span>
<span class="d-inline-block" tabindex="0" data-toggle="tooltip" title="Majority of work done or significant contribution by MAIL/SAIL members!">
<abbr class="badge bg-danger"><i class="bi bi-check2"></i></abbr>
</span>
One Loss for Quantization: Deep Hashing with Discrete Wasserstein Distributional Matching
</div>
<div class="author">
<u>Khoa D Doan</u>,
<a href="https://sites.google.com/site/pengyangshomepage/" target="_blank">Peng Yang</a>,
and <a href="http://research.baidu.com/People/index-view?id=111" target="_blank">Ping Li</a>
</div>
<div class="periodical">
<em>In Conference on Computer Vision and Pattern Recognition</em>
2022
</div>
<div class="links">
<a class="abstract btn btn-sm z-depth-0" role="button">Abs</a>
<a class="bibtex btn btn-sm z-depth-0" role="button">Bib</a>
<a href="https://openreview.net/pdf?id=uaqweIZ-9_k" class="btn btn-sm z-depth-0" role="button" target="_blank">PDF</a>
<a href="https://github.com/khoadoan106/single_loss_quantization" class="btn btn-sm z-depth-0" role="button" target="_blank">Code</a>
<a href="/assets/pdf/doan2022hswd-slides.pdf" class="btn btn-sm z-depth-0" role="button" target="_blank">Slides</a>
<a class="submissions btn btn-sm z-depth-0" role="button">Submission History</a>
</div>
<!-- Hidden abstract block -->
<div class="abstract hidden">
<p>Image hashing is a principled approximate nearest neighbor approach to find similar items to a query in a large collection of images. Hashing aims to learn a binary-output function that maps an image to a binary vector. For optimal retrieval performance, producing balanced hash codes with low-quantization error to bridge the gap between the learning stage's continuous relaxation and the inference stage's discrete quantization is important. However, in the existing deep supervised hashing methods, coding balance and low-quantization error are difficult to achieve and involve several losses. We argue that this is because the existing quantization approaches in these methods are heuristically constructed and not effective to achieve these objectives. This paper considers an alternative approach to learning the quantization constraints. The task of learning balanced codes with low quantization error is re-formulated as matching the learned distribution of the continuous codes to a pre-defined discrete, uniform distribution. This is equivalent to minimizing the distance between two distributions. We then propose a computationally efficient distributional distance by leveraging the discrete property of the hash functions. This distributional distance is a valid distance and enjoys lower time and sample complexities. The proposed single-loss quantization objective can be integrated into any existing supervised hashing method to improve code balance and quantization error. Experiments confirm that the proposed approach substantially improves the performance of several representative hashing methods.</p>
</div>
<!-- Hidden bibtex block -->
<div class="bibtex hidden">
<figure class="highlight"><pre><code class="language-bibtex" data-lang="bibtex"><span class="nc">@inproceedings</span><span class="p">{</span><span class="nl">doan2022hswd</span><span class="p">,</span>
<span class="na">bibtex_show</span> <span class="p">=</span> <span class="s">{true}</span><span class="p">,</span>
<span class="na">abbr</span> <span class="p">=</span> <span class="s">{CVPR}</span><span class="p">,</span>
<span class="na">title</span> <span class="p">=</span> <span class="s">{One Loss for Quantization: Deep Hashing with Discrete Wasserstein Distributional Matching}</span><span class="p">,</span>
<span class="na">author</span> <span class="p">=</span> <span class="s">{Doan, Khoa D and Yang, Peng and Li, Ping}</span><span class="p">,</span>
<span class="na">booktitle</span> <span class="p">=</span> <span class="s">{Conference on Computer Vision and Pattern Recognition}</span><span class="p">,</span>
<span class="na">year</span> <span class="p">=</span> <span class="s">{2022}</span><span class="p">,</span>
<span class="na">submissions</span> <span class="p">=</span> <span class="s">{CVPR'22}</span><span class="p">,</span>
<span class="na">slides</span> <span class="p">=</span> <span class="s">{doan2022hswd-slides.pdf}</span><span class="p">,</span>
<span class="na">code</span> <span class="p">=</span> <span class="s">{https://github.com/khoadoan106/single_loss_quantization}</span><span class="p">,</span>
<span class="na">pdf</span> <span class="p">=</span> <span class="s">{https://openreview.net/pdf?id=uaqweIZ-9_k}</span><span class="p">,</span>
<span class="na">teaser</span> <span class="p">=</span> <span class="s">{doan2022hswd.png}</span><span class="p">,</span>
<span class="na">selected</span> <span class="p">=</span> <span class="s">{true}</span>
<span class="p">}</span></code></pre></figure>
</div>
<!-- Hidden bibtex block -->
<div class="submissions hidden">
<p>CVPR'22</p>
</div>
</div>
</div>
</li>
<li>
<div class="row">
<!-- <div class="col-sm-1 abbr">
<div>
<abbr class="badge">NeurIPS</abbr>
</div>
</div> -->
<div id="doan2021wb" class="col-sm-12">
<div class="title">
<span class="abbr">
<abbr class="badge">NeurIPS</abbr>
</span>
<span class="d-inline-block" tabindex="0" data-toggle="tooltip" title="Majority of work done or significant contribution by MAIL/SAIL members!">
<abbr class="badge bg-danger"><i class="bi bi-check2"></i></abbr>
</span>
Backdoor Attack with Imperceptible Input and Latent Modification
</div>
<div class="author">
<u>Khoa D Doan</u>,
<a href="http://ylao.people.clemson.edu/" target="_blank">Yingjie Lao</a>,
and <a href="http://research.baidu.com/People/index-view?id=111" target="_blank">Ping Li</a>
</div>
<div class="periodical">
<em>In Thirty-Fifth Conference on Neural Information Processing Systems</em>
2021
</div>
<div class="links">
<a class="abstract btn btn-sm z-depth-0" role="button">Abs</a>
<a class="bibtex btn btn-sm z-depth-0" role="button">Bib</a>
<a href="https://proceedings.neurips.cc/paper/2021/file/9d99197e2ebf03fc388d09f1e94af89b-Paper.pdf" class="btn btn-sm z-depth-0" role="button" target="_blank">PDF</a>
<a href="https://recorder-v3.slideslive.com/?share=51522&s=8af881c0-56e8-451e-865f-adb1e90e5471" class="btn btn-sm z-depth-0" role="button" target="_blank">VIDEO</a>
<a href="https://github.com/khoadoan106/backdoor_attacks" class="btn btn-sm z-depth-0" role="button" target="_blank">Code</a>
<a href="/assets/pdf/doan2021wb-slides.pdf" class="btn btn-sm z-depth-0" role="button" target="_blank">Slides</a>
<a class="submissions btn btn-sm z-depth-0" role="button">Submission History</a>
</div>
<!-- Hidden abstract block -->
<div class="abstract hidden">
<p>Recent studies have shown that deep neural networks (DNN) are vulnerable to various adversarial attacks. In particular, an adversary can inject a stealthy backdoor into a model such that the compromised model will behave normally without the presence of the trigger. Techniques for generating backdoor images that are visually imperceptible from clean images have also been developed recently, which further enhance the stealthiness of the backdoor attacks from the input space. Along with the development of attacks, defense against backdoor attacks is also evolving. Many existing countermeasures found that backdoor tends to leave tangible footprints in the latent or feature space, which can be utilized to mitigate backdoor attacks.In this paper, we extend the concept of imperceptible backdoor from the input space to the latent representation, which significantly improves the effectiveness against the existing defense mechanisms, especially those relying on the distinguishability between clean inputs and backdoor inputs in latent space. In the proposed framework, the trigger function will learn to manipulate the input by injecting imperceptible input noise while matching the latent representations of the clean and manipulated inputs via a Wasserstein-based regularization of the corresponding empirical distributions. We formulate such an objective as a non-convex and constrained optimization problem and solve the problem with an efficient stochastic alternating optimization procedure. We name the proposed backdoor attack as Wasserstein Backdoor (WB), which achieves a high attack success rate while being stealthy from both the input and latent spaces, as tested in several benchmark datasets, including MNIST, CIFAR10, GTSRB, and TinyImagenet.</p>
</div>
<!-- Hidden bibtex block -->
<div class="bibtex hidden">
<figure class="highlight"><pre><code class="language-bibtex" data-lang="bibtex"><span class="nc">@inproceedings</span><span class="p">{</span><span class="nl">doan2021wb</span><span class="p">,</span>
<span class="na">bibtex_show</span> <span class="p">=</span> <span class="s">{true}</span><span class="p">,</span>
<span class="na">abbr</span> <span class="p">=</span> <span class="s">{NeurIPS}</span><span class="p">,</span>
<span class="na">title</span> <span class="p">=</span> <span class="s">{Backdoor Attack with Imperceptible Input and Latent Modification}</span><span class="p">,</span>
<span class="na">author</span> <span class="p">=</span> <span class="s">{Doan, Khoa D and Lao, Yingjie and Li, Ping}</span><span class="p">,</span>
<span class="na">booktitle</span> <span class="p">=</span> <span class="s">{Thirty-Fifth Conference on Neural Information Processing Systems}</span><span class="p">,</span>
<span class="na">year</span> <span class="p">=</span> <span class="s">{2021}</span><span class="p">,</span>
<span class="na">submissions</span> <span class="p">=</span> <span class="s">{NeurIPS'21}</span><span class="p">,</span>
<span class="na">pdf</span> <span class="p">=</span> <span class="s">{https://proceedings.neurips.cc/paper/2021/file/9d99197e2ebf03fc388d09f1e94af89b-Paper.pdf}</span><span class="p">,</span>
<span class="na">code</span> <span class="p">=</span> <span class="s">{https://github.com/khoadoan106/backdoor_attacks}</span><span class="p">,</span>
<span class="na">video</span> <span class="p">=</span> <span class="s">{https://recorder-v3.slideslive.com/?share=51522&s=8af881c0-56e8-451e-865f-adb1e90e5471}</span><span class="p">,</span>
<span class="na">slides</span> <span class="p">=</span> <span class="s">{doan2021wb-slides.pdf}</span><span class="p">,</span>
<span class="na">teaser</span> <span class="p">=</span> <span class="s">{doan2021wb.png}</span><span class="p">,</span>
<span class="na">selected</span> <span class="p">=</span> <span class="s">{true}</span>
<span class="p">}</span></code></pre></figure>
</div>
<!-- Hidden bibtex block -->
<div class="submissions hidden">
<p>NeurIPS'21</p>
</div>
</div>
</div>
</li>
<li>
<div class="row">
<!-- <div class="col-sm-1 abbr">
<div>
<abbr class="badge">ICCV</abbr>
</div>
</div> -->
<div id="doan2021lira" class="col-sm-12">
<div class="title">
<span class="abbr">
<abbr class="badge">ICCV</abbr>
</span>
<span class="d-inline-block" tabindex="0" data-toggle="tooltip" title="Majority of work done or significant contribution by MAIL/SAIL members!">
<abbr class="badge bg-danger"><i class="bi bi-check2"></i></abbr>
</span>
LIRA: Learnable, Imperceptible and Robust Backdoor Attacks
</div>
<div class="author">
<u>Khoa D Doan</u>,
<a href="http://ylao.people.clemson.edu/" target="_blank">Yingjie Lao</a>,
<a href="https://www.cs.rit.edu/~wjz/" target="_blank">Weijie Zhao</a>,
and <a href="http://research.baidu.com/People/index-view?id=111" target="_blank">Ping Li</a>
</div>
<div class="periodical">
<em>In International Conference on Computer Vision</em>
2021
</div>
<div class="links">
<a class="abstract btn btn-sm z-depth-0" role="button">Abs</a>
<a class="bibtex btn btn-sm z-depth-0" role="button">Bib</a>
<a href="https://openaccess.thecvf.com/content/ICCV2021/papers/Doan_LIRA_Learnable_Imperceptible_and_Robust_Backdoor_Attacks_ICCV_2021_paper.pdf" class="btn btn-sm z-depth-0" role="button" target="_blank">PDF</a>
<a href="https://github.com/khoadoan106/backdoor_attacks" class="btn btn-sm z-depth-0" role="button" target="_blank">Code</a>
<a href="https://github.com/sunbelbd/invisible_backdoor_attacks/raw/master/resources/ICCV2021-LIRA-Poster.pdf" class="btn btn-sm z-depth-0" role="button" target="_blank">Poster</a>
<a href="https://github.com/sunbelbd/invisible_backdoor_attacks/raw/master/resources/ICCV2021-LIRA-Slides.pdf" class="btn btn-sm z-depth-0" role="button" target="_blank">Slides</a>
<a class="submissions btn btn-sm z-depth-0" role="button">Submission History</a>
</div>
<!-- Hidden abstract block -->
<div class="abstract hidden">
<p>Recently, machine learning models have demonstrated to be vulnerable to backdoor attacks, primarily due to the lack of transparency in black-box models such as deep neural networks. A third-party model can be poisoned such that it works adequately in normal conditions but behaves maliciously on samples with specific trigger patterns. However, the trigger injection function is manually defined in most existing backdoor attack methods, e.g., placing a small patch of pixels on an image or slightly deforming the image before poisoning the model. This results in a two-stage approach with a sub-optimal attack success rate and a lack of complete stealthiness under human inspection.
In this paper, we propose a novel and stealthy backdoor attack framework, LIRA, which jointly learns the optimal, stealthy trigger injection function and poisons the model. We formulate such an objective as a non-convex, constrained optimization problem. Under this optimization framework, the trigger generator function will learn to manipulate the input with imperceptible noise to preserve the model performance on the clean data and maximize the attack success rate on the poisoned data. Then, we solve this challenging optimization problem with an efficient, two-stage stochastic optimization procedure. Finally, the proposed attack framework achieves 100% success rates in several benchmark datasets, including MNIST, CIFAR10, GTSRB, and T-ImageNet, while simultaneously bypassing existing backdoor defense methods and human inspection.
</p>
</div>
<!-- Hidden bibtex block -->
<div class="bibtex hidden">
<figure class="highlight"><pre><code class="language-bibtex" data-lang="bibtex"><span class="nc">@inproceedings</span><span class="p">{</span><span class="nl">doan2021lira</span><span class="p">,</span>
<span class="na">bibtex_show</span> <span class="p">=</span> <span class="s">{true}</span><span class="p">,</span>
<span class="na">abbr</span> <span class="p">=</span> <span class="s">{ICCV}</span><span class="p">,</span>
<span class="na">title</span> <span class="p">=</span> <span class="s">{LIRA: Learnable, Imperceptible and Robust Backdoor Attacks}</span><span class="p">,</span>
<span class="na">author</span> <span class="p">=</span> <span class="s">{Doan, Khoa D and Lao, Yingjie and Zhao, Weijie and Li, Ping}</span><span class="p">,</span>
<span class="na">booktitle</span> <span class="p">=</span> <span class="s">{International Conference on Computer Vision}</span><span class="p">,</span>
<span class="na">submissions</span> <span class="p">=</span> <span class="s">{ICCV'21}</span><span class="p">,</span>
<span class="na">code</span> <span class="p">=</span> <span class="s">{https://github.com/khoadoan106/backdoor_attacks}</span><span class="p">,</span>
<span class="na">slides</span> <span class="p">=</span> <span class="s">{https://github.com/sunbelbd/invisible_backdoor_attacks/raw/master/resources/ICCV2021-LIRA-Slides.pdf}</span><span class="p">,</span>
<span class="na">poster</span> <span class="p">=</span> <span class="s">{https://github.com/sunbelbd/invisible_backdoor_attacks/raw/master/resources/ICCV2021-LIRA-Poster.pdf}</span><span class="p">,</span>
<span class="na">pdf</span> <span class="p">=</span> <span class="s">{https://openaccess.thecvf.com/content/ICCV2021/papers/Doan_LIRA_Learnable_Imperceptible_and_Robust_Backdoor_Attacks_ICCV_2021_paper.pdf}</span><span class="p">,</span>
<span class="na">year</span> <span class="p">=</span> <span class="s">{2021}</span><span class="p">,</span>
<span class="na">teaser</span> <span class="p">=</span> <span class="s">{doan2021lira.png}</span><span class="p">,</span>
<span class="na">selected</span> <span class="p">=</span> <span class="s">{true}</span>
<span class="p">}</span></code></pre></figure>
</div>
<!-- Hidden bibtex block -->
<div class="submissions hidden">
<p>ICCV'21</p>
</div>
</div>
</div>
</li>
<li>
<div class="row">
<!-- <div class="col-sm-1 abbr">
<div>
<abbr class="badge">SIGIR</abbr>
</div>
</div> -->
<div id="doan2021interpretable" class="col-sm-12">
<div class="title">
<span class="abbr">
<abbr class="badge">SIGIR</abbr>
</span>
<span class="d-inline-block" tabindex="0" data-toggle="tooltip" title="Majority of work done or significant contribution by MAIL/SAIL members!">
<abbr class="badge bg-danger"><i class="bi bi-check2"></i></abbr>
</span>
Interpretable Graph Similarity Computation via Differentiable Optimal Alignment of Node Embeddings
</div>
<div class="author">
<u>Khoa D Doan</u>,
<a href="https://gurdaspuriya.github.io/" target="_blank">Saurav Manchanda</a>,
Suchismit Mahapatra,
and <a href="https://people.cs.vt.edu/reddy/index.html" target="_blank">Chandan K Reddy</a>
</div>
<div class="periodical">
<em>In 44th International ACM SIGIR Conference on Research and Development in Information Retrieval</em>
2021
</div>
<div class="links">
<a class="abstract btn btn-sm z-depth-0" role="button">Abs</a>
<a class="bibtex btn btn-sm z-depth-0" role="button">Bib</a>
<a href="/assets/pdf/doan2021interpretable.pdf" class="btn btn-sm z-depth-0" role="button" target="_blank">PDF</a>
<a href="https://www.youtube.com/watch?v=IWxxsuFPsgs&t=1s" class="btn btn-sm z-depth-0" role="button" target="_blank">VIDEO</a>
<a href="https://github.com/khoadoan/GraphOTSim" class="btn btn-sm z-depth-0" role="button" target="_blank">Code</a>
<a href="https://github.com/khoadoan/GraphOTSim/raw/main/resources/SIGIR21-fp0937-slides.pdf" class="btn btn-sm z-depth-0" role="button" target="_blank">Slides</a>
<a class="submissions btn btn-sm z-depth-0" role="button">Submission History</a>
</div>
<!-- Hidden abstract block -->
<div class="abstract hidden">
<p>Computing graph similarity is an important task in many graph-related applications
such as retrieval in graph databases or graph clustering. While numerous measures
have been proposed to capture the similarity between a pair of graphs, Graph Edit
Distance (GED) and Maximum Common Subgraphs (MCS) are the two widely used measures
in practice. GED and MCS are domain-agnostic measures of structural similarity between
the graphs and define the similarity as a function of pairwise alignment of different
entities (such as nodes, edges, and subgraphs) in the two graphs. The explicit explainability
offered by the pairwise alignment provides transparency and justification of the similarity
score, thus, GED and MCS have important practical applications. However, their exact
computations are known to be NP-hard. While recently proposed neural-network based
approximations have been shown to accurately compute these similarity scores, they
have limited ability in providing comprehensive explanations compared to classical
combinatorial algorithms, e.g., Beam search. This paper aims at efficiently approximating
these domain-agnostic similarity measures through a neural network, and simultaneously
learning the alignments (i.e., explanations) similar to those of classical intractable
methods. Specifically, we formulate the similarity between a pair of graphs as the
minimal "transformation" cost from one graph to another in the learnable node-embedding
space. We show that, if node embedding is able to capture its neighborhood context
closely, our proposed similarity function closely approximates both the alignment
and the similarity score of classical methods. Furthermore, we also propose an efficient
differentiable computation of our proposed objective for model training. Empirically,
we demonstrate that the proposed method achieves up to 50%-100% reduction in the Mean
Squared Error for the graph similarity approximation task and up to 20% improvement
in the retrieval evaluation metrics for the graph retrieval task. The source code
is available at https://github.com/khoadoan/GraphOTSim.</p>
</div>
<!-- Hidden bibtex block -->
<div class="bibtex hidden">
<figure class="highlight"><pre><code class="language-bibtex" data-lang="bibtex"><span class="nc">@inproceedings</span><span class="p">{</span><span class="nl">doan2021interpretable</span><span class="p">,</span>
<span class="na">bibtex_show</span> <span class="p">=</span> <span class="s">{true}</span><span class="p">,</span>
<span class="na">abbr</span> <span class="p">=</span> <span class="s">{SIGIR}</span><span class="p">,</span>
<span class="na">url</span> <span class="p">=</span> <span class="s">{https://doi.org/10.1145/3404835.3462960}</span><span class="p">,</span>
<span class="na">code</span> <span class="p">=</span> <span class="s">{https://github.com/khoadoan/GraphOTSim}</span><span class="p">,</span>
<span class="na">pdf</span> <span class="p">=</span> <span class="s">{doan2021interpretable.pdf}</span><span class="p">,</span>
<span class="na">slides</span> <span class="p">=</span> <span class="s">{https://github.com/khoadoan/GraphOTSim/raw/main/resources/SIGIR21-fp0937-slides.pdf}</span><span class="p">,</span>
<span class="na">video</span> <span class="p">=</span> <span class="s">{https://www.youtube.com/watch?v=IWxxsuFPsgs&t=1s}</span><span class="p">,</span>
<span class="na">teaser</span> <span class="p">=</span> <span class="s">{doan2021interpretable.png}</span><span class="p">,</span>
<span class="na">selected</span> <span class="p">=</span> <span class="s">{true}</span><span class="p">,</span>
<span class="na">submissions</span> <span class="p">=</span> <span class="s">{AAAI'21 -- WWW'21 -- SIGIR'21}</span><span class="p">,</span>
<span class="na">author</span> <span class="p">=</span> <span class="s">{Doan, Khoa D and Manchanda, Saurav and Mahapatra, Suchismit and Reddy, Chandan K}</span><span class="p">,</span>
<span class="na">title</span> <span class="p">=</span> <span class="s">{Interpretable Graph Similarity Computation via Differentiable Optimal Alignment of Node Embeddings}</span><span class="p">,</span>
<span class="na">year</span> <span class="p">=</span> <span class="s">{2021}</span><span class="p">,</span>
<span class="na">isbn</span> <span class="p">=</span> <span class="s">{9781450380379}</span><span class="p">,</span>
<span class="na">publisher</span> <span class="p">=</span> <span class="s">{Association for Computing Machinery}</span><span class="p">,</span>
<span class="na">address</span> <span class="p">=</span> <span class="s">{New York, NY, USA}</span><span class="p">,</span>
<span class="na">doi</span> <span class="p">=</span> <span class="s">{10.1145/3404835.3462960}</span><span class="p">,</span>
<span class="na">booktitle</span> <span class="p">=</span> <span class="s">{44th International ACM SIGIR Conference on Research and Development in Information Retrieval}</span><span class="p">,</span>
<span class="na">pages</span> <span class="p">=</span> <span class="s">{665–674}</span><span class="p">,</span>
<span class="na">numpages</span> <span class="p">=</span> <span class="s">{10}</span><span class="p">,</span>
<span class="na">keywords</span> <span class="p">=</span> <span class="s">{similarity search, model interpretability, graph similarity, GCN}</span><span class="p">,</span>
<span class="na">location</span> <span class="p">=</span> <span class="s">{Virtual Event, Canada}</span><span class="p">,</span>
<span class="na">series</span> <span class="p">=</span> <span class="s">{SIGIR '21}</span>
<span class="p">}</span></code></pre></figure>
</div>
<!-- Hidden bibtex block -->
<div class="submissions hidden">
<p>AAAI'21 -- WWW'21 -- SIGIR'21</p>
</div>
</div>
</div>
</li>
<li>
<div class="row">
<!-- <div class="col-sm-1 abbr">
<div>
<abbr class="badge">WWW</abbr>
</div>
</div> -->
<div id="doan2020efficient" class="col-sm-12">
<div class="title">
<span class="abbr">
<abbr class="badge">WWW</abbr>
</span>
<span class="d-inline-block" tabindex="0" data-toggle="tooltip" title="Majority of work done or significant contribution by MAIL/SAIL members!">
<abbr class="badge bg-danger"><i class="bi bi-check2"></i></abbr>
</span>
Efficient Implicit Unsupervised Text Hashing Using Adversarial Autoencoder
</div>
<div class="author">
<u>Khoa D Doan</u>,
and <a href="https://people.cs.vt.edu/reddy/index.html" target="_blank">Chandan K Reddy</a>
</div>
<div class="periodical">
<em>In Proceedings of The Web Conference</em>
2020
</div>
<div class="links">
<a class="abstract btn btn-sm z-depth-0" role="button">Abs</a>
<a class="bibtex btn btn-sm z-depth-0" role="button">Bib</a>
<a href="https://dl.acm.org/doi/abs/10.1145/3366423.3380150" class="btn btn-sm z-depth-0" role="button" target="_blank">HTML</a>
<a href="https://people.cs.vt.edu/~reddy/papers/WWW20a.pdf" class="btn btn-sm z-depth-0" role="button" target="_blank">PDF</a>
</div>
<!-- Hidden abstract block -->
<div class="abstract hidden">
<p>Searching for documents with semantically similar content is a fundamental problem
in the information retrieval domain with various challenges, primarily, in terms of
efficiency and effectiveness. Despite the promise of modeling structured dependencies
in documents, several existing text hashing methods lack an efficient mechanism to
incorporate such vital information. Additionally, the desired characteristics of an
ideal hash function, such as robustness to noise, low quantization error and bit balance/uncorrelation,
are not effectively learned with existing methods. This is because of the requirement
to either tune additional hyper-parameters or optimize these heuristically and explicitly
constructed cost functions. In this paper, we propose a Denoising Adversarial Binary
Autoencoder (DABA) model which presents a novel representation learning framework
that captures structured representation of text documents in the learned hash function.
Also, adversarial training provides an alternative direction to implicitly learn a
hash function that captures all the desired characteristics of an ideal hash function.
Essentially, DABA adopts a novel single-optimization adversarial training procedure
that minimizes the Wasserstein distance in its primal domain to regularize the encoder’s
output of either a recurrent neural network or a convolutional autoencoder. We empirically
demonstrate the effectiveness of our proposed method in capturing the intrinsic semantic
manifold of the related documents. The proposed method outperforms the current state-of-the-art
shallow and deep unsupervised hashing methods for the document retrieval task on several
prominent document collections.</p>
</div>
<!-- Hidden bibtex block -->
<div class="bibtex hidden">
<figure class="highlight"><pre><code class="language-bibtex" data-lang="bibtex"><span class="nc">@inproceedings</span><span class="p">{</span><span class="nl">doan2020efficient</span><span class="p">,</span>
<span class="na">bibtex_show</span> <span class="p">=</span> <span class="s">{true}</span><span class="p">,</span>
<span class="na">abbr</span> <span class="p">=</span> <span class="s">{WWW}</span><span class="p">,</span>
<span class="na">url</span> <span class="p">=</span> <span class="s">{https://doi.org/10.1145/3366423.3380150}</span><span class="p">,</span>
<span class="na">html</span> <span class="p">=</span> <span class="s">{https://dl.acm.org/doi/abs/10.1145/3366423.3380150}</span><span class="p">,</span>
<span class="na">pdf</span> <span class="p">=</span> <span class="s">{https://people.cs.vt.edu/~reddy/papers/WWW20a.pdf}</span><span class="p">,</span>
<span class="na">teaser</span> <span class="p">=</span> <span class="s">{doan2020efficient.png}</span><span class="p">,</span>
<span class="na">selected</span> <span class="p">=</span> <span class="s">{true}</span><span class="p">,</span>
<span class="na">author</span> <span class="p">=</span> <span class="s">{Doan, Khoa D and Reddy, Chandan K}</span><span class="p">,</span>
<span class="na">title</span> <span class="p">=</span> <span class="s">{Efficient Implicit Unsupervised Text Hashing Using Adversarial Autoencoder}</span><span class="p">,</span>
<span class="na">year</span> <span class="p">=</span> <span class="s">{2020}</span><span class="p">,</span>
<span class="na">isbn</span> <span class="p">=</span> <span class="s">{9781450370233}</span><span class="p">,</span>
<span class="na">publisher</span> <span class="p">=</span> <span class="s">{Association for Computing Machinery}</span><span class="p">,</span>
<span class="na">address</span> <span class="p">=</span> <span class="s">{New York, NY, USA}</span><span class="p">,</span>
<span class="na">doi</span> <span class="p">=</span> <span class="s">{10.1145/3366423.3380150}</span><span class="p">,</span>
<span class="na">booktitle</span> <span class="p">=</span> <span class="s">{Proceedings of The Web Conference}</span><span class="p">,</span>
<span class="na">pages</span> <span class="p">=</span> <span class="s">{684–694}</span><span class="p">,</span>
<span class="na">numpages</span> <span class="p">=</span> <span class="s">{11}</span><span class="p">,</span>
<span class="na">keywords</span> <span class="p">=</span> <span class="s">{autoencoder, Hashing, adversarial training, deep learning.}</span><span class="p">,</span>
<span class="na">location</span> <span class="p">=</span> <span class="s">{Taipei, Taiwan}</span><span class="p">,</span>
<span class="na">series</span> <span class="p">=</span> <span class="s">{WWW '20}</span>
<span class="p">}</span></code></pre></figure>
</div>
<!-- Hidden bibtex block -->
</div>
</div>
</li>
<li>
<div class="row">
<!-- <div class="col-sm-1 abbr">
<div>
<abbr class="badge">arXiv</abbr>
</div>
</div> -->
<div id="badirli2020gradient" class="col-sm-12">
<div class="title">
<span class="abbr">
<abbr class="badge">arXiv</abbr>
</span>
Gradient boosting neural networks: Grownet
</div>
<div class="author">
<a href="https://scholar.google.com/citations?user=Mkaq4VYAAAAJ&hl=en" target="_blank">Sarkhan Badirli</a>,
<a href="https://xuanqing94.github.io/" target="_blank">Xuanqing Liu</a>,
<a href="https://zmxing.github.io/" target="_blank">Zhengming Xing</a>,
<a href="https://avradeep1.github.io/" target="_blank">Avradeep Bhowmik</a>,
<u>Khoa D Doan</u>,
and <a href="http://www.keerthis.com/" target="_blank">Sathiya K Selvaraj</a>
</div>
<div class="periodical">
<em>arXiv preprint arXiv:2002.07971</em>
2020
</div>
<div class="links">
<a class="abstract btn btn-sm z-depth-0" role="button">Abs</a>
<a class="bibtex btn btn-sm z-depth-0" role="button">Bib</a>
<a href="https://arxiv.org/pdf/2002.07971.pdf" class="btn btn-sm z-depth-0" role="button" target="_blank">PDF</a>
<a href="https://github.com/sbadirli/GrowNet" class="btn btn-sm z-depth-0" role="button" target="_blank">Code</a>
</div>
<!-- Hidden abstract block -->
<div class="abstract hidden">
<p>A novel gradient boosting framework is proposed where shallow neural networks are employed as ``weak learners''. General loss functions are considered under this unified framework with specific examples presented for classification, regression, and learning to rank. A fully corrective step is incorporated to remedy the pitfall of greedy function approximation of classic gradient boosting decision tree. The proposed model rendered outperforming results against state-of-the-art boosting methods in all three tasks on multiple datasets. An ablation study is performed to shed light on the effect of each model components and model hyperparameters.</p>
</div>
<!-- Hidden bibtex block -->
<div class="bibtex hidden">
<figure class="highlight"><pre><code class="language-bibtex" data-lang="bibtex"><span class="nc">@article</span><span class="p">{</span><span class="nl">badirli2020gradient</span><span class="p">,</span>
<span class="na">bibtex_show</span> <span class="p">=</span> <span class="s">{true}</span><span class="p">,</span>
<span class="na">abbr</span> <span class="p">=</span> <span class="s">{arXiv}</span><span class="p">,</span>
<span class="na">url</span> <span class="p">=</span> <span class="s">{https://arxiv.org/abs/2002.07971}</span><span class="p">,</span>
<span class="na">code</span> <span class="p">=</span> <span class="s">{https://github.com/sbadirli/GrowNet}</span><span class="p">,</span>
<span class="na">pdf</span> <span class="p">=</span> <span class="s">{https://arxiv.org/pdf/2002.07971.pdf}</span><span class="p">,</span>
<span class="na">teaser</span> <span class="p">=</span> <span class="s">{badirli2020gradient.png}</span><span class="p">,</span>
<span class="na">selected</span> <span class="p">=</span> <span class="s">{true}</span><span class="p">,</span>
<span class="na">title</span> <span class="p">=</span> <span class="s">{Gradient boosting neural networks: Grownet}</span><span class="p">,</span>
<span class="na">author</span> <span class="p">=</span> <span class="s">{Badirli, Sarkhan and Liu, Xuanqing and Xing, Zhengming and Bhowmik, Avradeep and Doan, Khoa D and Selvaraj, Sathiya K}</span><span class="p">,</span>
<span class="na">journal</span> <span class="p">=</span> <span class="s">{arXiv preprint arXiv:2002.07971}</span><span class="p">,</span>
<span class="na">year</span> <span class="p">=</span> <span class="s">{2020}</span>
<span class="p">}</span></code></pre></figure>
</div>
<!-- Hidden bibtex block -->
</div>
</div>
</li>
<li>
<div class="row">
<!-- <div class="col-sm-1 abbr">
<div>
<abbr class="badge">CIKM</abbr>
</div>
</div> -->
<div id="doan2019adversarial" class="col-sm-12">
<div class="title">
<span class="abbr">
<abbr class="badge">CIKM</abbr>
</span>
<span class="d-inline-block" tabindex="0" data-toggle="tooltip" title="Majority of work done or significant contribution by MAIL/SAIL members!">
<abbr class="badge bg-danger"><i class="bi bi-check2"></i></abbr>
</span>
Adversarial Factorization Autoencoder for Look-Alike Modeling
</div>
<div class="author">
<u>Khoa D Doan</u>,
Pranjul Yadav,
and <a href="https://people.cs.vt.edu/reddy/index.html" target="_blank">Chandan K Reddy</a>
</div>
<div class="periodical">
<em>In Proceedings of the 28th ACM International Conference on Information and Knowledge Management</em>
2019
</div>
<div class="links">
<a class="abstract btn btn-sm z-depth-0" role="button">Abs</a>
<a class="bibtex btn btn-sm z-depth-0" role="button">Bib</a>
<a href="https://dmkd.cs.vt.edu/papers/CIKM19.pdf" class="btn btn-sm z-depth-0" role="button" target="_blank">PDF</a>
</div>
<!-- Hidden abstract block -->
<div class="abstract hidden">
<p>Digital advertising is performed in multiple ways, for e.g., contextual, display-based
and search-based advertising. Across these avenues, the primary goal of the advertiser
is to maximize the return on investment. To realize this, the advertiser often aims
to target the advertisements towards a targeted set of audience as this set has a
high likelihood to respond positively towards the advertisements. One such form of
tailored and personalized, targeted advertising is known as look-alike modeling, where
the advertiser provides a set of seed users and expects the machine learning model
to identify a new set of users such that the newly identified set is similar to the
seed-set with respect to the online purchasing activity. Existing look-alike modeling
techniques (i.e., similarity-based and regression-based) suffer from serious limitations
due to the implicit constraints induced during modeling. In addition, the high-dimensional
and sparse nature of the advertising data increases the complexity. To overcome these
limitations, in this paper, we propose a novel Adversarial Factorization Autoencoder
that can efficiently learn a binary mapping from sparse, high-dimensional data to
a binary address space through the use of an adversarial training procedure. We demonstrate
the effectiveness of our proposed approach on a dataset obtained from a real-world
setting and also systematically compare the performance of our proposed approach with
existing look-alike modeling baselines.</p>
</div>
<!-- Hidden bibtex block -->
<div class="bibtex hidden">
<figure class="highlight"><pre><code class="language-bibtex" data-lang="bibtex"><span class="nc">@inproceedings</span><span class="p">{</span><span class="nl">doan2019adversarial</span><span class="p">,</span>
<span class="na">bibtex_show</span> <span class="p">=</span> <span class="s">{true}</span><span class="p">,</span>
<span class="na">abbr</span> <span class="p">=</span> <span class="s">{CIKM}</span><span class="p">,</span>
<span class="na">pdf</span> <span class="p">=</span> <span class="s">{https://dmkd.cs.vt.edu/papers/CIKM19.pdf}</span><span class="p">,</span>
<span class="na">teaser</span> <span class="p">=</span> <span class="s">{doan2019adversarial.png}</span><span class="p">,</span>
<span class="na">selected</span> <span class="p">=</span> <span class="s">{true}</span><span class="p">,</span>
<span class="na">author</span> <span class="p">=</span> <span class="s">{Doan, Khoa D and Yadav, Pranjul and Reddy, Chandan K}</span><span class="p">,</span>
<span class="na">title</span> <span class="p">=</span> <span class="s">{Adversarial Factorization Autoencoder for Look-Alike Modeling}</span><span class="p">,</span>
<span class="na">year</span> <span class="p">=</span> <span class="s">{2019}</span><span class="p">,</span>
<span class="na">isbn</span> <span class="p">=</span> <span class="s">{9781450369763}</span><span class="p">,</span>
<span class="na">publisher</span> <span class="p">=</span> <span class="s">{Association for Computing Machinery}</span><span class="p">,</span>
<span class="na">address</span> <span class="p">=</span> <span class="s">{New York, NY, USA}</span><span class="p">,</span>
<span class="na">url</span> <span class="p">=</span> <span class="s">{https://doi.org/10.1145/3357384.3357807}</span><span class="p">,</span>
<span class="na">doi</span> <span class="p">=</span> <span class="s">{10.1145/3357384.3357807}</span><span class="p">,</span>
<span class="na">booktitle</span> <span class="p">=</span> <span class="s">{Proceedings of the 28th ACM International Conference on Information and Knowledge Management}</span><span class="p">,</span>
<span class="na">pages</span> <span class="p">=</span> <span class="s">{2803–2812}</span><span class="p">,</span>
<span class="na">numpages</span> <span class="p">=</span> <span class="s">{10}</span><span class="p">,</span>
<span class="na">keywords</span> <span class="p">=</span> <span class="s">{deep learning, autoencoder, hashing, factorization, adversarial training, look-alike modeling}</span><span class="p">,</span>
<span class="na">location</span> <span class="p">=</span> <span class="s">{Beijing, China}</span><span class="p">,</span>
<span class="na">series</span> <span class="p">=</span> <span class="s">{CIKM '19}</span>
<span class="p">}</span></code></pre></figure>
</div>
<!-- Hidden bibtex block -->
</div>
</div>
</li>
</ol>
<div>
<a class="btn btn-sm" role="button" style="color: black; border: 1px solid black; font-style: normal; padding: 0.0rem 0.0rem 0.0rem 0.0rem;">Submission History</a> shows the venues where the work has been submitted (🙃 including <strong>rejections</strong> 🙃). I hope some of my poor rejection/failure histories (record now is 10 rejections 😅) give you some encouragement to try again when things don't work out (don't give up -- good work doesn't need to be rushed)!
</div>
</div>
<hr>
<div class="news">
<h2>Open Office Hour</h2>
<p>
I will ocassionally be holding group open office hours (fully ONLINE) for *anyone*.
Feel free to <a href="#open-office-hour-calendar">sign up</a> to connect, chat, or ask any questions.
</p>
<p>When I was a student, I was clueless sometimes (if not most of the time) and I had no idea how to get help.
I hope that, via this modest effort, I can share some experience with you, as well as address some questions you may have,
using my experience working in both industry and academia and applied and research projects,
as well as experience in studying abroad in the US.
<strong>I encourage to converse in English.</strong>
<!-- Calendly inline widget begin -->
<!-- <div id="open-office-hour-calendar" class="calendly-inline-widget" data-url="https://calendly.com/khoa-vinuni/mail-open-office-hour" style="min-width:320px;height:950px;"></div>
<script type="text/javascript" src="https://assets.calendly.com/assets/external/widget.js" async></script> -->
<!-- Calendly inline widget end -->
<!-- </div> -->
</p>
<div id="open-office-hour-calendar" class="calendly-inline-widget" style="position: relative;min-width:320px;height:550px;" data-processed="true">
<div class="calendly-spinner">
<div class="calendly-bounce1"></div>
<div class="calendly-bounce2"></div>
<div class="calendly-bounce3"></div>
</div>
<iframe src="https://outlook.office365.com/owa/calendar/MAILPublic@vinuni.edu.vn/bookings/s/wtWIaegmL0iFvgFOI-c-jA2" width="100%" height="100%" frameborder="0"></iframe>
</div>
<p>This effort is inspired by <a href="https://mlcollective.org/services/#mloh">ML Collective</a></p>
<!-- <script type='text/javascript' id='clustrmaps' src='//cdn.clustrmaps.com/map_v2.js?cl=ffffff&w=300&t=n&d=6Aazaa3W3ib-sAIfPEQYBd0A-PM2ZbkZfCUcYo54GsE'></script> -->
</div>
</article>
</div>
<!-- Footer -->
<div class="container mt-0">
<em>The brick walls are there for a reason. The brick walls are not there to keep us out. The brick walls are there to give us a chance to show how badly we want something</em> -- <a href="https://en.wikipedia.org/wiki/Randy_Pausch">Randy Pausch</a>
</div>
<script type="text/javascript" id="clustrmaps" src="//cdn.clustrmaps.com/map_v2.js?cl=ffffff&w=200&t=n&d=6Aazaa3W3ib-sAIfPEQYBd0A-PM2ZbkZfCUcYo54GsE&co=233d4d"></script>
<footer class="fixed-bottom">
<div class="container mt-0">
© Copyright 2026 Khoa D Doan.
</div>
</footer>
</div></body>
<!-- Bootsrap & MDB scripts -->
<script src="https://cdnjs.cloudflare.com/ajax/libs/popper.js/2.4.4/umd/popper.min.js" integrity="sha512-eUQ9hGdLjBjY3F41CScH3UX+4JDSI9zXeroz7hJ+RteoCaY+GP/LDoM8AO+Pt+DRFw3nXqsjh9Zsts8hnYv8/A==" crossorigin="anonymous"></script>
<script src="https://stackpath.bootstrapcdn.com/bootstrap/4.5.2/js/bootstrap.min.js" integrity="sha512-M5KW3ztuIICmVIhjSqXe01oV2bpe248gOxqmlcYrEzAvws7Pw3z6BK0iGbrwvdrUQUhi3eXgtxp5I8PDo9YfjQ==" crossorigin="anonymous"></script>
<script src="https://cdnjs.cloudflare.com/ajax/libs/mdbootstrap/4.19.1/js/mdb.min.js" integrity="sha512-Mug9KHKmroQFMLm93zGrjhibM2z2Obg9l6qFG2qKjXEXkMp/VDkI4uju9m4QKPjWSwQ6O2qzZEnJDEeCw0Blcw==" crossorigin="anonymous"></script>
<!-- Mansory & imagesLoaded -->
<script defer src="https://unpkg.com/masonry-layout@4/dist/masonry.pkgd.min.js"></script>
<script defer src="https://unpkg.com/imagesloaded@4/imagesloaded.pkgd.min.js"></script>
<script defer src="/assets/js/mansory.js" type="text/javascript"></script>
<!-- Medium Zoom JS -->
<script src="https://cdn.jsdelivr.net/npm/medium-zoom@1.0.6/dist/medium-zoom.min.js" integrity="sha256-EdPgYcPk/IIrw7FYeuJQexva49pVRZNmt3LculEr7zM=" crossorigin="anonymous"></script>
<script src="/assets/js/zoom.js"></script>
<!-- Load Common JS -->
<script src="/assets/js/common.js"></script>
</html>
Sitemap
Кол-во: 0
XML-карта сайта для поисковиков
?
Sitemap.xml помогает поисковику быстрее находить и индексировать страницы. Особенно важен для крупных сайтов и новых страниц, на которые ещё нет входящих ссылок.
Robots.txt не содержит ссылку на карту сайта. Рекомендуется добавить карту сайта и указать ссылку на нее в robots.txt.
Внутренние ссылки
Кол-во: 13
Ссылки на другие страницы своего сайта
?
Внутренние ссылки распределяют ссылочный вес между страницами и помогают поисковику обходить сайт. Пустые анкоры и ссылки на запрещённые robots.txt страницы — типичные ошибки.
Внутренних ссылок на странице 13 оптимально.
Внутренние ссылки не запрещены к индексации в robots.txt.
Показать внутренние ссылки
| Url | Анкор | Состояние | Анализировать |
|---|---|---|---|
| /research |
research
|
|
Анализировать url |
| /publications/ |
publications
|
|
Анализировать url |
| /team |
group
|
|
Анализировать url |
| /photos/ |
photos
|
|
Анализировать url |
| /teaching/ |
teaching
|
|
Анализировать url |
| /contact |
contact
|
|
Анализировать url |
| /research |
more about our research
|
|
Анализировать url |
| / |
[AAAI’26-a]
|
|
Анализировать url |
| /team/ |
Nguyen K Hoang
|
|
Анализировать url |
| /publications |
full list
|
|
Анализировать url |
| /assets/pdf/doan2022hswd-slides.pdf |
Slides
|
|
|
| /assets/pdf/doan2021wb-slides.pdf |
Slides
|
|
|
| /assets/pdf/doan2021interpretable.pdf |
PDF
|
|
Внешние ссылки
Кол-во: 151
Ссылки на сторонние сайты
?
Исходящие внешние ссылки передают часть ссылочного веса на чужие сайты. Ссылки на авторитетные ресурсы безопасны; ссылки на мусорные сайты могут навредить репутации страницы.
Внешних ссылок на странице 151 слишком много. Спрячьте лишние ссылки в тег noindex или атрибут rel='nofollow'!
Показать первые 100 внешних ссылок
| Url | Анкор | Анализировать |
|---|---|---|
| mail-research.com |
<img width="" height="30px" alt="blank" src="/assets/img/logo/logo-mail-v2.png">
|
Анализировать url |
| github.com |
<i class="fab fa-github" style="color: red;"></i>
|
Анализировать url |
| linkedin.com |
<i class="fab fa-linkedin"></i>
|
Анализировать url |
| twitter.com |
<i class="fab fa-twitter"></i>
|
|
| facebook.com |
<i class="fab fa-facebook"></i>
|
|
| sail-research.com |
SAIL Research
|
Анализировать url |
| khoahocphattrien.vn |
khoahocphattrien
|
Анализировать url |
| thanhnien.vn |
Thanh Nien
|
Анализировать url |
| vnexpress.net |
VnExpress
|
Анализировать url |
| baodautu.vn |
BaoDauTu
|
Анализировать url |
| dantri.com.vn |
DanTri
|
Анализировать url |
| vietnam.vn |
Vietnam.vn
|
Анализировать url |
| vietnam.vn |
Vietnam.vn
|
Анализировать url |
| sg.finance.yahoo.com |
Yahoo Finance
|
Анализировать url |
| benzinga.com |
Benzinga
|
Анализировать url |
| macaubusiness.com |
Macau Business
|
Анализировать url |
| taiwannews.com.tw |
Taiwan News
|
Анализировать url |
| technode.global |
TNGlobal
|
Анализировать url |
| vingroup.net |
VinGroup
|
Анализировать url |
| x.com |
Posts by @khoaddoan
|
Анализировать url |
| forms.gle |
here
|
Анализировать url |
| forms.gle |
here
|
Анализировать url |
| arxiv.org |
[COLM’26-a]
|
Анализировать url |
| arxiv.org |
[COLM’26-b]
|
Анализировать url |
| tridentatmm26mgc.github.io |
TRIDENT challenge
|
Анализировать url |
| csie.ntu.edu.tw |
Wen-Huang Cheng
|
Анализировать url |
| basiclab.lab.nycu.edu.tw |
Hong-Han Shuai
|
Анализировать url |
| hongxiaxie.net |
Hongxia Xie
|
Анализировать url |
| lynn-ling-lo.github.io |
Ling Lo
|
Анализировать url |
| github.com |
Jian-Yu Jiang-Lin
|
Анализировать url |
| alberthkyhky.github.io |
Kang-Yang Huang
|
Анализировать url |
| linkedin.com |
Ling Zou
|
Анализировать url |
| openreview.net |
[TMLR’26]
|
Анализировать url |
| openreview.net |
[ICLR’26]
|
Анализировать url |
| openreview.net |
[NeurIPS’25-a]
|
Анализировать url |
| openreview.net |
[NeurIPS’25-b]
|
Анализировать url |
| openreview.net |
[NeurIPS’25-c]
|
Анализировать url |
| openreview.net |
[NeurIPS-W’25-d]
|
Анализировать url |
| dl.acm.org |
ACM AI Letters
|
Анализировать url |
| mtuann.github.io |
Tuan M Nguyen
|
Анализировать url |
| arxiv.org |
[ICML’25-a]
|
Анализировать url |
| openreview.net |
[ICML-W’25-b]
|
Анализировать url |
| people.cs.vt.edu |
Chandan K Reddy
|
Анализировать url |
| arxiv.org |
PDF
|
Анализировать url |
| github.com |
Code
|
Анализировать url |
| hieunt91.github.io |
Hieu T Nguyen
|
Анализировать url |
| angli.ai |
Ang Li
|
Анализировать url |
| vietanhnguyen.net |
Viet Anh Nguyen
|
Анализировать url |
| openreview.net |
PDF
|
Анализировать url |
| github.com |
Code
|
Анализировать url |
| research.baidu.com |
Ping Li
|
Анализировать url |
| cs.rit.edu |
Weijie Zhao
|
Анализировать url |
| ylao.people.clemson.edu |
Yingjie Lao
|
Анализировать url |
| github.com |
Code
|
Анализировать url |
| arxiv.org |
PDF
|
Анализировать url |
| github.com |
Code
|
Анализировать url |
| arxiv.org |
PDF
|
Анализировать url |
| github.com |
Code
|
Анализировать url |
| people.cs.vt.edu |
Chandan K Reddy
|
Анализировать url |
| arxiv.org |
PDF
|
Анализировать url |
| github.com |
Code
|
Анализировать url |
| huggingface.co |
Data
|
Анализировать url |
| the-anhta.github.io |
The-Anh Ta
|
Анализировать url |
| thanhnguyentang.github.io |
Thanh Nguyen-Tang
|
Анализировать url |
| sail-research.com |
Kok-Seng Wong
|
Анализировать url |
| scholar.google.com |
Hoang Thanh-Tung
|
Анализировать url |
| openreview.net |
PDF
|
Анализировать url |
| github.com |
Code
|
Анализировать url |
| the-anhta.github.io |
The-Anh Ta
|
Анализировать url |
| arxiv.org |
PDF
|
Анализировать url |
| github.com |
Code
|
Анализировать url |
| docs.google.com |
Slides
|
Анализировать url |
| scholar.google.com.au |
Tung Pham
|
Анализировать url |
| ecva.net |
PDF
|
Анализировать url |
| gurdaspuriya.github.io |
Saurav Manchanda
|
Анализировать url |
| niteshchawla.nd.edu |
Nitesh V Chawla
|
Анализировать url |
| arxiv.org |
PDF
|
Анализировать url |
| github.com |
Code
|
Анализировать url |
| hieunt91.github.io |
Hieu Trung Nguyen
|
Анализировать url |
| duykhuongnguyen.github.io |
Duy Nguyen
|
Анализировать url |
| vietanhnguyen.net |
Viet Anh Nguyen
|
Анализировать url |
| arxiv.org |
PDF
|
Анализировать url |
| cs.cornell.edu |
Eugene Bagdasaryan
|
Анализировать url |
| sail-research.com |
Kok-Seng Wong
|
Анализировать url |
| scholar.google.com |
Anh T Tran
|
Анализировать url |
| cs-chan.com |
Chee Seng Chan
|
Анализировать url |
| arxiv.org |
PDF
|
Анализировать url |
| github.com |
Code
|
Анализировать url |
| gurdaspuriya.github.io |
Saurav Manchanda
|
Анализировать url |
| v-mipeng.github.io |
Minlong Peng
|
Анализировать url |
| sail-research.com |
Kok-Seng Wong
|
Анализировать url |
| arxiv.org |
PDF
|
Анализировать url |
| github.com |
Code
|
Анализировать url |
| ylao.people.clemson.edu |
Yingjie Lao
|
Анализировать url |
| scholar.google.com.au |
Tung Pham
|
Анализировать url |
| sail-research.com |
Kok-Seng Wong
|
Анализировать url |
| openreview.net |
PDF
|
Анализировать url |
| github.com |
Code
|
Анализировать url |
| scholar.google.com |
Anh T Tran
|
Анализировать url |
| sail-research.com |
Kok-Seng Wong
|
Анализировать url |
Конкуренты Готовность: 0%
Конкуренты в Яндексе
Кол-во: 0
Топ сайтов-конкурентов в Яндексе
?
Сайты, чаще всего появляющиеся в ТОПе Яндекса по запросам из семантического ядра этой страницы.
Мы не нашли у вас конкурентов в Яндексе. Сайт или очень молодой или плохо продвигается.
Конкурентов в ТОП-10 Яндекса не нашлось.
Конкуренты в Google
Кол-во: 0
Топ сайтов-конкурентов в Google
?
Сайты, чаще всего появляющиеся в ТОПе Google по запросам из семантического ядра этой страницы.
Конкуренты в Google тоже не найдены. Займитесь продвижением сайта!
Конкурентов в ТОП-10 Google не нашлось.
ЗоЗПП: права потребителей Готовность: 100%
Нарушения
Не выявлены
Признаков дистанционной продажи товаров (интернет-магазина) не обнаружено — требования ЗоЗПП о раскрытии информации продавца к сайту не применяются. Нарушений нет.
ФЗ-149: рекомендательные технологии Готовность: 100%
Нарушения
Не выявлены
Рекомендательные блоки («с этим покупают», «похожие товары» и т.п.) на сайте не обнаружены — требования ст. 10.7 ФЗ-149 к сайту не применяются. Нарушений нет.
ФЗ-38: реклама Готовность: 100%
Нарушения
Не выявлены
Рекламных тематик с обязательными оговорками (медицина, БАД, кредиты и займы, новостройки) на сайте не обнаружено. Нарушений нет.
ФЗ-436: защита детей Готовность: 100%
Нарушения
Не выявлены
Признаков информационной продукции (новости, видео, книги, игры, курсы) не обнаружено — обязательная возрастная маркировка по ФЗ-436 сайту не требуется. Нарушений нет.
Вердикт
Оптимизация сайта khoadoan.me выполнена всего на 36%. Чтобы исправить ситуацию и попасть в ТОП необходимо:
Исправьте ошибки в мета-тегах.
Исправьте ошибки оптимизации.
Исправьте ошибки индексации.
Поделитесь с друзьями: