Qi Pang
Ph.D. Candidate, Computer Science Department, Carnegie Mellon University
I am a Ph.D. candidate in the Computer Science Department at Carnegie Mellon University, advised by Prof. Wenting Zheng and Prof. Virginia Smith.
I am broadly interested in privacy and safety in foundation models. My research focuses on (1) building efficient, trustworthy AI systems with provable cryptographic guarantees, and (2) uncovering unknown threats to agentic systems before they lead to real breaches, with the ultimate goal of making AI systems more reliable.
I am currently on the job market.
Email: qipang [at] cmu.edu Google Scholar GitHub
Research
I use cryptography and ML insights to both make provable guarantees practical for modern AI systems and uncover attacks that evade today's defenses.
- Building trustworthy AI systems with provable guarantees. Our work enables efficient private Transformer inference (BOLT, FABLE), proposes trustworthy synthetic text generation that needs no fine-tuning (Guided Diffusion), finds and repairs cases where privacy protection silently changes a model's predictions (MPCDiff), and lets model trainers efficiently prove that they applied the promised privacy mechanisms (Noisette).
- Uncovering new vulnerabilities. Our recent research shows that LLM agents can collude through normal-looking messages that monitoring provably cannot detect (Codetta), and that the same properties that make LLM watermarks successful also let attackers remove or spoof them (No Free Lunch).
Complementary research on federated and distributed learning
- I've done complementary research on robust and privacy-preserving federated and distributed learning, including attacks on vertical federated learning (ADI), Byzantine-robust federated learning with near-optimal statistical rates, and secure federated statistics enabling non-linear statistics aggregation.
Publications (Google Scholar)
* equal contribution
Preprints
-
Codetta: High-Capacity, Keyless, and Undetectable Multi-Agent CollusionQi Pang, Virginia Smith, Wenting Zheng
-
Guided Diffusion for Trustworthy Synthetic Text GenerationQi Pang, Hanshen Xiao, Andre Manoel, Lipika Ramaswamy, Wenting Zheng, Virginia Smith, G. Edward SuhUnder review, 2026
Conference Papers
-
Noisette: Certifying Differential Privacy Mechanisms EfficientlyQi Pang, Radhika Garg, Ziling Liu, Hanshen Xiao, Virginia Smith, Wenting Zheng, Xiao WangIEEE S&P 2027 (to appear) [Paper]
-
FABLE: Batched Evaluation on Confidential Lookup Tables in 2PCZhengyuan Su, Qi Pang, Simon Beyzerov, Wenting Zheng
-
BOLT: Privacy-Preserving, Accurate and Efficient Inference for TransformersQi Pang, Jinhao Zhu, Helen Möllering, Wenting Zheng, Thomas Schneider
-
No Free Lunch in LLM Watermarking: Trade-offs in Watermarking Design ChoicesQi Pang, Shengyuan Hu, Wenting Zheng, Virginia Smith
-
Communication Bounds for the Distributed Experts ProblemZhihao Jia, Qi Pang, Trung Tran, David Woodruff, Zhihao Zhang, Wenting Zheng (alphabetical order)
-
MPCDiff: Testing and Repairing MPC-Hardened Deep Learning ModelsQi Pang, Yuanyuan Yuan, Shuai Wang
-
ADI: Adversarial Dominating Inputs in Vertical Federated Learning SystemsQi Pang, Yuanyuan Yuan, Shuai Wang, Wenting Zheng
-
Secure Federated Correlation Test and Entropy EstimationQi Pang*, Lun Wang*, Shuai Wang, Wenting Zheng, Dawn Song
-
Byzantine-Robust Federated Learning with Optimal Statistical RatesBanghua Zhu*, Lun Wang*, Qi Pang*, Shuai Wang, Jiantao Jiao, Dawn Song, Michael I. Jordan
Earlier publications (2020–2022)
-
MDPFuzz: Testing Models Solving Markov Decision ProcessesQi Pang, Yuanyuan Yuan, Shuai Wang
-
PrivGuard: Privacy Regulation Compliance Made EasierLun Wang, Usmann Khan, Joseph Near, Qi Pang, Jithendaraa Subramanian, Neel Somani, Peng Gao, Andrew Low, Dawn Song
-
Automated Side Channel Analysis of Media Software with Manifold LearningYuanyuan Yuan, Qi Pang, Shuai Wang
-
mID: Tracing Screen Photos via Moiré PatternsYushi Cheng, Xiaoyu Ji, Lixu Wang*, Qi Pang*, Yi-Chao Chen, Wenyuan XuUSENIX Security 2021 [Paper]
-
Towards Practical Differentially Private Causal Graph DiscoveryLun Wang, Qi Pang, Dawn Song
Talks
-
High-Capacity, Keyless, and Undetectable Multi-Agent Collusion
Stanford Security Seminar; Berkeley Security Seminar; CMU Crypto Seminar2026 -
FABLE: Batched Evaluation on Confidential Lookup Tables in 2PC
USENIX Security2025 -
No Free Lunch in LLM Watermarking: Trade-offs in Watermarking Design Choices
Amazon Research Crypto Seminar (2025); CMU CyLab Partners Conference (2024)2024–2025 -
BOLT: Privacy-Preserving, Accurate and Efficient Inference for Transformers
IEEE S&P; Google Research GenAI PSS Seminar; Berkeley Security Seminar2024 -
MPCDiff: Testing and Repairing MPC-Hardened Deep Learning Models
NDSS2024 -
ADI: Adversarial Dominating Inputs in Vertical Federated Learning Systems
IEEE S&P2023
Education & Experience
-
Carnegie Mellon University, Ph.D. in Computer Science
Advisors: Wenting Zheng and Virginia Smith2022 – present -
NVIDIA Research, Research Intern
Mentor: Hanshen Xiao; Manager: G. Edward SuhMay 2025 – May 2026 -
HKUST, M.Phil. in Computer Science and Engineering
Advisor: Shuai Wang2020 – 2022 -
UC Berkeley, Research Intern
Advisor: Dawn SongNov. 2019 – Aug. 2022 -
Zhejiang University, B.Eng. in Electronic Information Engineering
Advanced Honor Class of Engineering Education (ACEE), Chu Kochen Honors College
Advisors: Wenyuan Xu and Xiaoyu Ji2016 – 2020
Awards
- Qualcomm Innovation Fellowship Finalist2024
- CyLab Presidential Fellowship, Carnegie Mellon University2024
Teaching & Service
- Teaching assistant. 18-330 Introduction to Computer Security, CMU (Spring 2026); 10-718 Machine Learning in Practice, CMU (Spring 2025); COMP3632 Principles of Cybersecurity, HKUST (Spring 2021, Fall 2021).
- Guest lecturer. 18-330 Introduction to Computer Security, CMU (Spring 2026); 15-793 Secure Computer Systems, CMU (Spring 2025).
- Program committee. USENIX Security 2027.
- Reviewer. ICML 2021–2026; NeurIPS 2021–2026.