Abstract
Copyright protection of the Federated Learning (FL) model has become a major concern since malicious clients in FL can stealthily distribute or sell the FL model to other parties. In order to prevent such misbehavior, one must be able to catch the culprit by investigating trace evidence from the model in question. In this paper, we propose FedTracker, the first FL model protection framework that, on one hand, employs global watermarks to verify ownerships of the global model; and on the other hand, embed unique local fingerprints into respective local models to facilitate tracing the model back to the culprit. Furthermore, FedTracker introduces the intuition of Continual Learning (CL) into watermark embedding, and proposes a CL-based watermark mechanism to improve fidelity. Experimental results show that the proposed FedTracker is effective in ownership verification, traceability, fidelity, and robustness.
Abstract (translated)
URL
https://arxiv.org/abs/2211.07160
