EmbTracker: a traceable black-box watermark for federated language models

EmbTracker is a server-side traceable black-box watermarking framework for federated language models (FedLM); by embedding a backdoor-based watermark, it can be verified with simple API queries. It injects a unique identity watermark into the model distributed to each client, enabling client-level tracing and attributing a leaked model to a specific client without participant cooperation. On multiple language and vision-language models, the verification rate is close to 100%; it resists removal attacks such as fine-tuning, pruning and quantization, and its impact on main-task performance is usually within 1-2%.