Overview

AI systems are increasingly deployed in safety-critical settings, from autonomous vehicles to software development tools. Our research studies how to make these systems more robust, how to detect and repair their failures, and how to understand the security risks introduced by the AI toolchain itself.

This spans work on DNN testing, adversarial robustness, autonomous driving simulation, compiler-level attacks on ML models, and causal approaches to system configuration.

Key Directions

• DNN Testing & Repair: Systematic methods for finding confusion and bias errors in neural networks, and techniques for repairing them.
• Autonomous Driving: Simulation-based testing, traffic generation, and scenario-based evaluation for self-driving systems.
• Adversarial Robustness: Metric learning and multitask approaches that strengthen model robustness under adversarial and natural variations.
• ML Toolchain Security: Discovering and exploiting vulnerabilities in deep learning compilers and infrastructure.
• Causal Configuration: Using causal reasoning to understand and optimize the performance of configurable systems.

Impact

DeepTest was a pioneering work in automated testing of autonomous driving systems and has been highly cited. Our work on compiler backdoors was accepted at IEEE S&P 2026, revealing a new attack surface in the ML supply chain. The Unicorn and Cameo systems brought causal reasoning to system performance optimization.

Contributors

Selected Publications