WaterMoE: Efficient, High-Fidelity Watermarking for MoE LLMs

A new watermarking technique called WaterMoE has been developed to address the growing need for content provenance and misuse detection in large language models (LLMs), particularly for the increasingly popular Mixture-of-Experts (MoE) architectures. Traditional watermarking methods often suffer from significant performance degradation and introduce considerable inference overhead, making them impractical for real-world deployment, especially in systems where latency is critical. WaterMoE tackles these issues by embedding watermarking signals directly into the expert selection process within each router of an MoE model. This controlled perturbation accumulates to subtly shift token selection at the final output. Unlike conventional watermarking approaches that typically act as a post-processing step for token sampling, WaterMoE integrates the watermark within the core inference loop. This innovative integration results in negligible quality degradation of the generated content and minimal computational overhead. Extensive experiments demonstrate that WaterMoE achieves fidelity performance nearly identical to unwatermarked models and consistently surpasses state-of-the-art watermarking methods across various generation tasks. It offers up to a 4x speedup and adds only about 1% additional inference latency compared to native generation, proving its viability for deployment in demanding, real-world applications.