�鶹APP

The College of Engineering, Applied Science, and Health presents its ECE seminar series featuring Xu Zhang, a Ph.D. student in computer engineering at Illinois Tech, who will present “AI Safety Under Inference-Time Manipulation: From Adversarial Robustness to Jailbreak Robustness.” This seminar will take place on Wednesday, October 28, from 12:45–1:45 p.m. in the Siegal Hall Auditorium (room 118).

Abstract

AI models can perform well on standard inputs yet behave unexpectedly when their inputs are manipulated by an adversary. Understanding the mechanisms behind these failures is essential to designing effective defenses. This talk examines that challenge in two settings: adversarial robustness in mixture-of-experts (MoE) classifiers and jailbreak robustness in multimodal large language models. In the first setting, we investigate the vulnerabilities of routers and experts, then develop expert-level robustification and joint training strategies to improve robustness and accuracy. In the second, we study how harmful in-context demonstrations steer model behavior, using a posterior-reweighting framework to explain observed attack trends and motivate a risk-gated defense based on benign counterevidence. Together, these studies illustrate a mechanism-guided approach to AI safety: diagnose vulnerabilities, explain behavioral changes, and design targeted interventions while preserving useful performance.

Biography

 Xu Zhang is a Ph.D. student in computer engineering at Illinois Tech, advised by Assitant Professor Ren Wang. His research focuses on AI safety and robust machine learning, including adversarial robustness in mixture-of-experts models and jailbreak robustness in multimodal large language models.