Research Statement

Building robust, secure, and trustworthy machine learning systems for real-world deployment.

Security through obscurity is not security. Guided by this principle, my work emphasizes adversarial evaluation under transparent, adaptive, and realistic threat models. I develop methods for evaluating and improving the robustness of machine learning systems under adversarial attacks, distribution shifts, corruptions, and real-world constraints. My research is grounded in adversarial machine learning and robust optimization, with applications to safety-critical perception, autonomous-driving perception, multimodal models, and emerging foundation-model systems.

A central theme of my research is adversarial evaluation: designing threat models, attacks, and stress tests that reveal failure modes missed by standard benchmarks. My past work introduced methods such as PDPGD, TsallisPGD, and ℂ-APGD for stronger robustness evaluation across norm-bounded, multi-norm, convex-hull, and semantic-segmentation settings. I have also developed defenses including Adversary Critic, Interpolated Adversarial Training, robust federated-learning aggregation, and moving-target defenses for deployed perception systems.

My research agenda is structured around three core themes:

• Adversarial Robustness & Robust Optimization — Developing principled objectives, optimization methods, and architectures for robustness under distribution shift, corruption, and adversarial attack.
• Real-World Robustness for Perception Systems — Designing physically grounded stress-testing methods and defenses for perception systems and world models under realistic deployment conditions.
• Security of Foundation Models — Extending adversarial and robustness evaluation to multimodal and foundation models, focusing on multimodal inputs, inference-time defenses, poisoning, and reliability failures.

Looking forward, I aim to extend these directions toward multimodal and agentic systems, where robustness failures arise from cross-modal interactions, tool use, and deployment feedback loops. My long-term goal is to establish formal and empirical foundations for trustworthy AI, enabling verifiable safety guarantees and the reliable deployment of robust and secure systems in safety-critical applications, including autonomous systems and healthcare.