I'm interested in AI4testing and Autonomous Driving testing. Most of my research is about detecting the anomalies targeting agents. Some projects are highlighted.
This project aims to detect and explain attack-induced anomalies in web applications by learning normal behavioral during runtime. We model the normal constraints with first-order logic, and generate executable Python scripts. By leveraging LLMs in the learning process, this approach enhances both the detection and explanation of anomalies.
This project proposes a novel and stealthy backdoor attack framework targeting deep learning-based lane detection systems in autonomous driving. Unlike previous methods that rely on random trigger placement and visually obvious patterns, this project strategically identifies high-sensitivity regions using gradient-based attention heatmaps to guide optimal trigger placement. It then leverages a diffusion-based generation pipeline to synthesize natural-looking triggers (e.g., cones or mud) that blend seamlessly into the scene. To ensure visual coherence and stealthiness, This project incorporates two loss functions that preserve lane structure and environmental consistency. Extensive experiments show that this project significantly outperforms existing attacks, achieving higher attack success rates while remaining nearly undetectable by forensic tools or human inspection.