Verifiable Homomorphic Secret Sharing for Machine Learning Classifiers

Abstract

When using machine learning classifiers to classify data in cloud computing, it is crucial to maintain data privacy and ensure the correctness of classification results. To address these security concerns, we propose a new verifiable homomorphic secret sharing (VHSS) scheme. Our approach involves distributing the task of executing a polynomial form of the machine learning classifier among two servers who produce partial results on encrypted data. Each server cannot obtain any data information, and the classification result can be reconstructed and verified using a verification key in conjunction with the two partial results. Compared to previous VHSS schemes, our scheme can compute the degree of polynomials as high as the polynomials in the system security parameters while performing comparably to homomorphic secret sharing (HSS) schemes. We implement our proposed scheme and demonstrate its application to decision trees (a type of machine learning classifier). Our experiments show that our scheme is twice as fast as previous VHSS schemes when evaluating decision trees with depths ranging from 2-14.