A Privacy-Preserving Distributed Machine Learning Protocol Based on Homomorphic Hash Authentication

Privacy-preserving machine learning is a hot topic in Artificial Intelligence (AI) area. However, there are also many security issues in all stages of privacy-oriented machine learning. This paper focuses on the dilemma that the privacy leakage of server-side parameter aggregation and external eavesdropper tampering during message transmission in the distributed machine learning framework. Combining with secret sharing techniques, we present a secure privacy-preserving distributed machine learning protocol under the double-server model based on homomorphic hash function, which enables our protocol verifiable. We also prove that our protocol can meet client semi-honest security requirements. Besides, we evaluate our protocol by comparing with other mainstream privacy preserving frameworks, in the aspects of computation, communication complexity analysis, in addition to a concrete implementation from the perspective of model convergence rate and execution time. Experimental results demonstrate that the local training model tends to converge at nearly 50 epochs where the convergence time is less than 400 s.