Targeting Federated Learning: A Study of Membership Inference Attacks on Healthcare Data

This study investigates the vulnerabilities of federated learning models in the healthcare domain, specifically focusing on membership inference attacks (MIA). Federated learning allows local models to train on sensitive healthcare data without sharing the data itself, making it an attractive method for protecting privacy. However, even in this decentralized framework, models remain vulnerable to MIAs, where attackers can infer whether certain data points were used to train a model by analyzing model updates. Using the Texas100 dataset, this study demonstrates that as the number of local models increases, the attack accuracy of MIAs also increases due to higher bias variations and model specialization. The proposed MIA model operates in a white-box setting, comparing global model updates with local model biases to identify the origin of sensitive data contributions from local participants. The findings underscore the need for stronger privacy protections in federated learning, particularly when applied to sensitive healthcare data.