Byzantine-Resilient Federated Learning: Evaluating MPC Approaches

Federated learning (FL) has emerged as a paradigm shift for collaborative machine learning to preserve data privacy. However, without considering the security measures through relevant cryptographic mechanisms, the collaborative process is vulnerable to various attacks. This paper evaluates the strength and scalability of Semi2k protocol for secure Multi Party Computation (MPC) under two major attacks, namely label-flipping and min-max attacks. We established a controlled simulations involving various numbers of malicious clients and MPC nodes. Our result showed that Semi2k offers limited protection against min-max attacks, showing no advantage over non-MPC setups in short training runs. However, it significantly improves accuracy under label-flipping attacks at 500 iterations, though overall accuracy declines with more malicious clients. Longer training improves resilience to label-flipping but increases communication overhead. Communication costs grow linearly with participants, highlighting a trade-off between scalability and efficiency.