InfoResearchPreprintLLM-specific
Plug-and-Play Multi-Licensee Robust Watermarking for Mixture-of-Experts LLMs
- Published
- Record updated
Summary
PPMoE is a watermarking framework for Mixture-of-Experts LLMs that lets a model owner embed a licensee-specific watermark without retraining. It trains mergeable incremental weight modules that steer the MoE router toward distinct routing patterns, then merges unique combinations into the base model at distribution time. Experiments across five MoE models and seven benchmark datasets report very high attribution accuracy, preserved base model utility, and robustness to adaptive watermark-removal attacks.