Lei Yu is a Tenure-Track Assistant Professor in the Department of Computer Science at Rensselaer Polytechnic Institute. His research focuses on data privacy, AI security, and trustworthy machine learning. He studies privacy risks across the data and model lifecycle, from membership inference in deep learning and fine-tuned language models to privacy-preserving federated learning. He also builds principled algorithmic and system-level defenses that protect privacy while keeping data and models useful.
He leads the Data Security and Privacy Lab (DSPLab) at RPI. The lab’s code and artifacts are on GitHub, and you can meet the team on the People page.
Before joining RPI, he was a Research Staff Member at the IBM T. J. Watson Research Center. There he worked on privacy protection for diagnostic system data, log-based anomaly detection, AIOps, and ML system optimization. He received his Ph.D. in Computer Science from the Georgia Institute of Technology, where he worked on big-data and deep-learning privacy. He earlier did doctoral work on wireless sensor networks at Harbin Institute of Technology.
Recent News
- A preprint on “Activation-Conditioned Self-Distillation” is now available on arXiv. PDF
- A preprint on membership inference against on-policy distillation, “Leaky Students: Membership Inference against On-Policy Distillation”, is now available on arXiv. PDF
- Our paper “SHE-LoRA: Selective Homomorphic Encryption for Federated Tuning with Heterogeneous LoRA” has been accepted to ICLR 2026. PDF
- Our paper “In-Context Probing for Membership Inference in Fine-Tuned Language Models” has been accepted to NDSS 2026. Congratulations to Zhexi! PDF Code
- A preprint on selective homomorphic encryption for joint federated learning in cross-device scenarios is now available on arXiv. PDF
- A preprint on the dynamics of membership privacy in deep learning is now available on arXiv. PDF
- Our paper “On the Adversarial Robustness of Graph Neural Networks with Graph Reduction” has been accepted to ESORICS 2025. Congratulations to our undergraduate researcher Kerui Wu! PDF Code
- Our paper “Membership Inference Attacks as Privacy Tools: Reliability, Disparity and Ensemble” has been accepted to CCS 2025. Congratulations to our undergraduate researchers Zhiqi Wang (first author), Chengyu Zhang, and Yuetian Chen! PDF Code
Selected Publications
- In-Context Probing for Membership Inference in Fine-Tuned Language ModelsNDSS 2026PaperPDFCode
- SHE-LoRA: Selective Homomorphic Encryption for Federated Tuning with Heterogeneous LoRAICLR 2026PaperPDF
- Membership Inference Attacks as Privacy Tools: Reliability, Disparity and EnsembleCCS 2025PaperPDFCode
- CMASan: Custom Memory Allocator-aware Address SanitizerIEEE S&P 2025PaperCode
- Privacy and Accuracy-Aware AI/ML Model DeduplicationSIGMOD 2025PaperPDF