Office: SEIEE Building, No.3, 526A
Dr. LIN, Yun (林云)
-
My research lies in automatic programming, explainable AI, and web/AI security.
My research is to compile complex requirements (especially on its complex knowledge system) into executable software systems, design and infer actionable explanation for LLMs and agents, and defend cybercrime such as phishing/scam and prompt injection attacks on LLMs.
- Automatic Programming:
- Requirement Compilation (code repository generation/editing): Compile and evolve complex specification into runnable software projects (ISSTA'26, OOPSLA'26, ASE'25, ISSTA'24, NeurIPS'23, EMNLP'22, FSE'15);
- Software Testing: Generate test cases for requirement and branch coverage (ISSTA'25, FSE'23, TDSC'22, ESEC/FSE'21, ISSTA'20, ICSE'20, ICSE'18);
- Software Debugging: Debugging process synthesis and root cause analysis (ICSE'26, ASE'24, ISSTA'24, ISSTA'22, ICSE'17, ASE'18, TSE'19, FSE'19);
- Explainable AI (and Its Applications): representation interpretation, training data attribution, and their applications.
- Web and AI Security:
- Web Misinformation: Detecting and explaining fake websites, emails, and SMSs, and monitoring the world phishing campaigns (CCS'26, USENIX Sec'25, USENIX Sec'24a, USENIX Sec'24b, USENIX Sec'23, USENIX Sec'22, USENIX Sec'21).
- AI Security: Defending prompt injection attacks on LLMs (CCS'26);
Research Works
Agentic Requirement Compilation (compiling spec to executables)
We build tools to compile and evolve complex specification into runnable software projects in a test-driven way. In those works, we consider specifications as executable constraints and compile them into software tests and formal verification models to guide the repository-wise code generation. See more details can be referred to Agentic Requirement Compiler (ISSTA'26), IntentionTest (ISSTA'26), IssueExec (ISSTA'26), TestGeneralizer (FSE'26), WebTestPilot (FSE'26).
Training Analytics (Explinable AI) and Its Applications
We build interactive and visualized approach for training models with explainable insights, e.g., how the model is trained and how the training data affects the model prediction.
Some work can be referred to XSearch (ISSTA'26)，TrainRef (ICLR'26)， DeepVisualInsight (AAAI'22), Empirical Influence Function (NeurIPS'22).
Web Misinformation and Agent Safety
We detect fake website, email, and SMS with agentic counterfactual analysis. We also focus on training models and building agents to defend prompt injection attacks on LLMs, which is a new security threat in the AIGC era.
Some work can be referred to DRIP (CCS'26), PiMRef (CCS'26), Halligan (USENIX Sec'25), PhishLLM (USENIX Sec'24）, PhishDecloaker (USENIX Sec'24), DynaPhish (USENIX Sec'23), PhishIntention (USENIX Sec'22), Phishpedia (USENIX Sec'21).