Biography: Prof. Jun Dai is an Associate Professor in the Department of Computer Science at Worcester Polytechnic Institute (WPI). He received his Ph.D. degree in Information Sciences and Technology from Pennsylvania State University, with a specialization in cybersecurity. He also holds a master’s degree in Network Science and Engineering and a bachelor’s degree in Information Security from University of Science and Technology of China. His research spans networks, distributed systems, artificial intelligence, and cybersecurity, with recent focus on large language model security, autonomous agent security, advanced attack detection, vulnerability analysis, secure coding, and cybersecurity education. Dr. Dai has published extensively in leading venues, including NDSS, ICML, ACM SenSys, ACM SIGMOD, IEEE TDSC, IEEE TIFS, and ACM SIGCSE. He currently serves as an Associate Editor for IEEE TDSC and regularly reviews for top-tier conferences such as ACM CCS and ICDCS, as well as premier journals including TIFS, TDSC, TVT, and TMC. He served as Workshop Chair for CCS 2023 and is currently co-chairing the Artifact Evaluation Committee for CCS 2026.
Speech Title: From Internal Representations to Trustworthy LLMs: Robustness, Privacy, and Provable Defense
Abstract: As Large Language Models (LLMs) become foundational to scientific discovery, data-driven decision making, and autonomous systems, ensuring their trustworthiness has emerged as a central challenge. The integration of external knowledge through Retrieval-Augmented Generation (RAG) and the widespread adoption of open-source LLMs significantly expand both capability and risk, introducing new attack surfaces and raising fundamental questions about robustness, privacy, and reliability. This talk advances a representation-centric perspective on LLM security, arguing that the key to trustworthy AI lies not only in what models produce, but in how they internally represent knowledge. By analyzing neural activations across layers, it becomes possible to detect adversarial manipulation, characterize the imprint of training data, and design systems that are resilient by construction. Building on this insight, the talk traces a progression from activation-based detection of poisoned responses, to understanding data memorization through internal representations, and ultimately to the development of provably robust retrieval mechanisms that bound adversarial influence. Together, these results point toward a new paradigm for securing AI systems, shifting from reactive, output-level safeguards to principled, representation-level reasoning, and outline a path toward trustworthy AI systems capable of operating reliably in adversarial, data-rich, and high-stakes environments.
Biography: Kaizhu Huang works on trustworthy machine learning, and neural/biomedical information processing. Before joining DKU, he was a full professor at Xi’an Jiaotong-Liverpool University (XJTLU) and Associate Dean of research in School of Advanced Technology, XJTLU. He was also Head of EEE, at XJTLU from 2016 to 2019. He worked at Fujitsu Research Centre, CUHK, University of Bristol, National Laboratory of Pattern Recognition, Chinese Academy of Sciences from 2004 to 2012. He was the recipient of the 2011 Asia Pacific Neural Network Society Young Researcher Award. He received the best paper or book awards for seven times. He has published 10 books and over 280 international research papers including 150+ journal papers (e.g. IEEE T-PAMI, IEEE T-IP, IEEE T-NNLS, IEEE T-CYB, JMLR) and 140+ conference papers (e.g. AAAI, IJCAI, SIGIR, NeurIPS, UAI, CIKM, ICDM, ICML, ECML, CVPR, ICCV). He is Editor in Chief, Elsevier CSSI and serves as associated editors/advisory board members in a number of international journals and book series. He was invited as a keynote speaker in more than 50 international conferences or workshops. He has led 5 NSFC major or general program projects, all as PI.
Biography: Prof. Zhi Li is a distinguished member of China Computer Federation (CCF), former standing committee member of its Technical Council on Software Engineering (TCSE), and a member of its Technical Council on Systems Software, Service Computing and Formal Methods, senior member of IEEE and ACM. He graduated with a BSc degree from Fudan University in 1991, an MSc degree from the University of York in 2004, and a PhD degree from The Open University in 2008. Prof. Li had spent over 10 years doing professional and technical work before he entered academia in 2001, with subsequent 9 years in the UK. His research interests are modeling, verifying, testing and validating Human-Cyber-Physical Systems (HCPSs) based on a problem-oriented approach, Artificial Intelligence for Requirements Engineering (AI4RE), Artificial General Intelligence (AGI), and Human-Computer Interaction (HCI). His research has been sponsored by 3 grants from the National Natural Science Foundation of China, and 5 grants from Ministry of Education of China, Guangxi Natural Science Foundation, and Guangxi Scientific Research & Technological Development. He has published over 80 research papers at international journals and conferences (including TSE, TKDE, FSE2025, Journal of Software, 3 best conference papers). He has given 2 keynote speeches and over 20 invited talks in international conferences, and he is the leader and one of the main contributors to a suite of Computer-Aided Requirements Engineering (CARE) tools for Problem-Oriented Software Development (POSD).
Copyright © 2026 the 9th International Conference on Big Data and Artificial Intelligence (BDAI 2026). All rights reserved.
