Speaker: 王安宇研究员，清华大学

Title: From PRC to CEC: Cryptanalysis and Construction of AI Watermarking Primitives

Language: Chinese

Time & Venue: 2026.06.16 10:00-11:00 南楼N402

Abstract: The rapid growth of AI-generated content calls for robust watermarking primitives that can reliably verify content origin. Pseudorandom Error-Correcting Codes (PRC) were recently introduced as the first primitive that provably achieves undetectability, robustness, and soundness simultaneously. In this talk, we present the first concrete security analysis of the LDPC-based PRC scheme used in state-of-the-art watermarking for LLMs and GIMs. Our cryptanalysis reveals that under practical parameters, the actual security of PRC drops significantly below the claimed level. We further identify a fundamental entropy bottleneck that renders PRC-based LLM watermarking impractical. Motivated by these findings, we introduce Confidential Error-Correcting Codes (CEC), a new primitive that relaxes undetectability to IND-CPA confidentiality while preserving robustness and soundness. We design a modular construction framework and instantiate it with strong error-correcting codes. The resulting watermarking scheme, CECMark, achieves 256-bit confidentiality, survives a wide range of image perturbations with >80% accuracy, and is deployable on current generative models.