A secure distributed machine learning protocol against static semi-honest adversaries

Abstract

Machine learning has been successfully applied to various fields over the last few years. However, it still faces two critical challenges. On one hand, the concern for the security issue in machine learning is increased. On the other hand, data exists in the form of isolated islands across different organizations. In this work, we focus on the privacy-preserving issue on a non-parametric machine learning algorithmize, i.e., the k Nearest Neighbor Classification (k NNC) in which, training data are split vertically among multiple servers. We propose a novel protocol that is secure against static semi-honest adversaries. In specific, the clients can obtain the label of his/her query without disclosing the servers’ data, the client’s query, and the client’s output to others. We use the state-of-the-art lattice-based fully homomorphic encryption to realize the privacy-preserving distance computation. In order to protect data access patterns, permutation technique and Oblivious Transfer are used in the top-k selection phase. We proved the security via the simulation paradigm. Meanwhile, we implemented our protocol and performed extensive experiments. Results show that our protocol performs well, especially in a large-width environment. Compared to the existing solution, our protocol leaks no information about the participants’ private input and output in both centralized and distributed architectures. Meanwhile, our protocol runs faster than existing solutions.