Periodic watermarking for copyright protection of large language models in cloud computing security

Pei-Gen Ye, Zhengdao Li, Zuopeng Yang, Pengyu Chen, Zhenxin Zhang, Ning Li, Jun Zheng

Abstract: Highlights•Innovative Temporal Watermarking: Introduces TimeMarker, a novel framework for embedding distinct watermarks across sub-periods to detect the timing of model extraction attacks on Large Language Models (LLMs).•Adaptive Watermark Strength: Utilizes adaptive watermark strength based on information entropy and frequency domain transformations, enhancing detection accuracy and robustness.•Broad Applicability: Validated across five widely used datasets, demonstrating effective detection of model extraction across various sub-periods and scenarios.•Enhanced EaaS Security: Extends traditional watermarking techniques, offering a robust solution for the copyright protection of Embedding as a Service (EaaS) models.•Contribution to AI Security: Provides insights into improving security standards and protecting intellectual property in AI-driven services.