Privacy Preservation and Attack Resistance in Blockchain-Enhanced Federated Learning for Medical Data

Abstract

Using federated learning (FL) in health care applications, the teams will be able to collaborate to predict the health of patients without exchanging any personally identifiable information. Regardless of this benefit, FL is still not widely used since it is susceptible to serious privacy and security risks like membership inference and model poisoning attacks. To address these concerns, this paper proposes a blockchain-enhanced FL model that involves safe aggregation and audit-trail immutability to thwart privacy leakage and manipulation of medical data analysis by opponents. Random Forest and LightGBM classifiers are trained on fake healthcare datasets in various scenarios of attack. Some of the security measures that are used to assess the system include attack success rate, cost of communication, differential privacy (DP) epsilon values and blockchain-based tamper detection delay. The proposed approach can achieve measurable privacy leaks reduction and a 40-percent lower success rate in membership inference attacks, as demonstrated in experiments. Also, blockchain audit trails can provide real-time resilience against tampering of data. According to these findings, blockchain-enhanced FL appears to be a promising framework of medical AI systems capable of ensuring the patient data safety and privacy.