Active Source Inference Attack Based on Label-Flipping in Federated Learning

Federated Learning (FL) is a distributed learning framework that creates a global model by periodically exchanging model parameters through a central server, without directly sharing local data. We introduce a novel Active Source Inference Attack (ASI) aimed at FL, which identifies the originating client of datasets within the FL system. ASI specifically includes a minor label-flipping operation on a small portion of data at the server side to train the global model, while concurrently using a supervised learning model to perform the inference attack. Extensive experiments on multiple datasets confirm the effectiveness of ASI, showing significant enhancements in attack success rate, recall, and precision. Additionally, ASI has minimal impact on the FL model, maintaining high attack stealth.