Seminars

View all Seminars  |  Download ICal for this event

Enhancing Safety in LLMs and other Foundation Models and Challenges with Enforcing Privacy Policies with LLMs  

Series: Department Seminar

Speaker: Varun Chandrasekaran, Assistant Professor, Department of Electrical and Computer Engineering, University of Illinois Urbana-Champaign and Affiliate Researcher, Microsoft Research

Date/Time: Oct 29 11:00:00

Location: CSA Auditorium, (Room No. 104, Ground Floor)

Abstract:
Foundation models are increasingly deployed in high-stakes environments, yet ensuring their safety remains a pressing challenge. This talk explores recent advancements in understanding and mitigating their risks, drawing on four key studies. We will examine (1) new frameworks for evaluating and aligning model behavior with human intent, (2) the security and reliability of watermarking techniques in foundation models, including their role in provenance tracking and their vulnerabilities to adversarial removal and evasion, and (3) novel approaches for detecting and mitigating high-risk model outputs before deployment. By synthesizing these findings, we will discuss the broader implications of foundation model security, trade-offs between robustness and control, and future directions for improving AI safety at scale.

Enforcing privacy policies in large-scale systems requires bridging a persistent gap between legal text and executable code. This talk explores how large language models (LLMs) can help and where they fall short across three critical stages of the pipeline. First, understanding policies is hard: privacy regulations are written in ambiguous natural language that even humans interpret inconsistently, let alone machines. Second, annotating policies is challenging: translating high-level requirements into structured, machine-checkable rules remains brittle, risking misalignment between policy intent and system implementation. Third, AI understanding of code is challenging: ultimately, policies must be enforced in code, but todays models struggle with program analysis, transformation, and invariant preservation. Drawing on recent projects, including work on LLM-based code translation, I argue that privacy compliance can be reframed as a problem of semantic preservation: just as we seek to preserve type and memory safety in translation, we must also preserve privacy invariants when policies are mapped into software. The talk will outline this three-part research agenda, highlight current limitations, and propose future directions toward building policy-aware, invariant-preserving AI systems.

Speaker Bio:
Varun Chandrasekaran is an Assistant Professor in the Department of Electrical and Computer Engineering at the University of Illinois Urbana-Champaign and an Affiliate Researcher at Microsoft Research. His research focuses on the intersection of security & privacy and AI/ML, with a recent emphasis on understanding and mitigating risks in foundation models. His work has been recognized with research awards from Amazon (2024), Microsoft Research (2024), and 2x Google (2025).

Host Faculty: Prof. Vinod Ganapathy