FedMCT: A Federated Framework for Intellectual Property Protection and Malicious Client Tracking

In the era of big data, federated learning (FL) emerges as a solution to train models collectively without exposing individual data, maintaining similar accuracy to models trained on shared datasets. However, challenges arise with the advent of privacy inference attacks and model theft, posing significant threats to the privacy of FL models, especially regarding intellectual property (IP) protection. This paper introduces FedMCT (Federated Malicious Client Tracking), a novel framework addressing these challenges in the FL context. The FedMCT framework is a new approach to protect IP rights of FL clients and track cheaters, which can improve efficiency in resource-heterogeneous environments. By embedding unique watermarks or fingerprints in Deep Neural Network (DNN) models, we can protect model IP. We employ a configuration round before watermark embedding, segmenting clients based on performance for tiered model watermarking. We also propose a tiered watermarking and traitor tracking mechanism, which reduces the tracking time and ensures high traitor tracking efficiency. Extensive experiments validate our solution's efficacy in maintaining original model performance, watermark privacy, and detectability, robust against various attacks, demonstrating superior traitor tracing efficiency compared to existing frameworks.