Blockchained Trustable Federated Learning Utilizing Voting Accountability for Malicious Actor Mitigation

Abstract

The federated learning (FL) approach in machine learning preserves user privacy during data collection. However, traditional FL schemes still rely on a centralized server, making them vulnerable to security risks, such as data breaches and tampering of models caused by malicious actors attempting to gain access by masquerading as trainers. To address these issues that hamper the trustability of federated learning, requirements were analyzed for several of these problems. The findings revealed that issues, such as the lack of accountability management, malicious actor mitigation, and model leakage, remained unaddressed in prior works. To fill this gap, a blockchain-based trustable FL scheme, MAM-FL, is proposed with the focus on providing accountability to trainers. MAM-FL established a group of voters responsible for evaluating and verifying the validity of the model updates submitted. The effectiveness of MAM-FL was tested based on the reduction of malicious actors present on both trainers’ and voters’ sides and the ability to handle colluding participants. Experiments show that MAM-FL succeeded at reducing the number of malicious actors, despite the test case involving initial collusion in the system.