A robust federated learning privacy framework: Integrating multi-party computation and advanced privacy-preserving techniques for secure data collaboration

Abstract

Federated Learning(FL) can train models without exposing raw client data, but still vulnerable to privacy leakage and poisoning attacks. Present methods like secure aggregation tend to offer little to no malicious update detection. Other research methods to mitigate this problem also fail to give a proper balance between privacy and robustness, or often give an impractical premise of needing multiple servers, creating a complicated application and deployment. This paper proposes DDFed-Markov, a probabilistic dual-defense federated learning framework using markovian chain that simultaneously improves privacy without compromising robustness. The framework uses a combination of noise injection using Markovian probability, Fully Homomorphic Encryption (FHE), and a feedback-driven strategy based on consensus with the collaboration of the Dual Defense Federated Learning Framework. While other models that use the concept of noise work in a similar way, the Markovian noise remains different because of its state-dependent nature. It introduces temporal correlation that increases resistance to adversaries while maintaining learning efficiency. Encrypted updates are aggregated using CKKS homomorphic encryption, allowing similarity-based aggregation without the need for multiple servers, preserving FLs’ hub and spoke topology. Also, client-side consensus is applied to filter malicious updates using adaptive thresholds. Experimental evaluations on publicly available data sets like MNIST have shown that DDFed-Markov has managed to achieve a comparative accuracy while effectively prevents potential model poisoning and successfully preserving privacy compared to existing frameworks.