Foundation models are increasingly deployed through prompts, fine-tuning, parameter-efficient adapters, retrieval, external tools, and autonomous agents. These mechanisms introduce new attack surfaces, including jailbreaks, backdoors, compromised model components, and malicious internal states.

Our research develops practical security and assurance mechanisms for these emerging AI ecosystems. Recent work includes dynamic defenses against jailbreak attacks, runtime detection of backdoored LoRA adapters, and analysis of how compromised AI components can affect downstream applications. More broadly, we seek continuous assurance methods that preserve model utility while improving the security of deployed foundation models and AI agents.

Selected Work:

• Dynamic Deep Prompt Optimization for Defending Against Jailbreak Attacks on LLMs
• LoRAScan: Detecting Backdoor Prompts in Low-Rank Adapters for Large Language Models via Down-Projection Activation Spikes
• CloudLeak: Large-Scale Deep Learning Models Stealing Through Adversarial Examples

As AI moves from digital environments into autonomous vehicles, robots, and other cyber-physical systems, security failures can propagate from sensing and perception to reasoning, planning, control, and ultimately physical action. Our research studies security across this complete embodied-AI loop.

Our work spans physical attacks against LiDAR and 3D perception, secure intelligent sensing, and emerging threats to LLM-powered robotic systems. Recent research shows that even a robot's own internal state and action history can become a hidden attack surface, motivating end-to-end defenses that jointly protect perception, reasoning, and physical execution.

Selected Work:

• Silent Sabotage: Internal State Triggered Backdoor Attacks on LLM-Powered Robotic Systems
• Robust Roadside Physical Adversarial Attack Against Deep Learning in LiDAR Perception Modules
• 3D-Adv: Black-Box Adversarial Attacks against Deep Learning Models through 3D Sensors
• Robust Adversarial Objects against Deep Learning Models

This research is supported in part by the NSF ERI project Towards Robust and Secure Intelligent 3D Sensing Systems (Award No. 2347426).

The trustworthiness of modern AI systems depends not only on models and data, but also on the software and hardware stacks on which they execute. Our research investigates security across these layers, including information leakage, side-channel analysis, hardware vulnerabilities, and interactions between AI models and their underlying computing platforms.

We also study the reverse direction: how AI and foundation models can support hardware design, debugging, and security assurance. Our recent work explores domain-specific and reasoning language models for hardware generation and revision, while our future work seeks to combine model behavior, software execution, and hardware-level signals for cross-layer AI assurance and secure hardware/software co-design.

Selected Work:

• Building Reasoning LLMs for Hardware Design Generation via Function-Aligned Differentiated Revision
• Hardware Phi-1.5B: A Large Language Model Encodes Hardware Domain Specific Knowledge
• LLM4SecHW: Leveraging Domain-Specific Large Language Model for Hardware Debugging
• DeepEM: Deep Neural Networks Model Recovery through EM Side-Channel Information Leakage