Privacy-preserving Machine Learning Based on Homomorphic Conjugate Search Problem

Abstract

With the development of computer technology and the rise of artificial intelligence, machine learning has achieved major breakthroughs in many fields. However, the accuracy of machine learning is usually proportional to the amount of training data, so a large amount of data needs to be collected. In the context of big data, both users and service providers have to bear the threaten of privacy disclosure since the attackers may abuse their data for illegal profits. On the other hand, it is necessary to prevent the leakage of model parameters in consideration of intellectual property protection. Therefore, it is imperative to solve privacy preservation for both data providers and trainers in machine learning. Without crippling the accuracy of machine learning, this paper proposed a neural network privacy training scheme in virtue of an efficient homomorphic encryption scheme to protect the privacy of data. The proposed scheme also realizes one-way security based on the Conjugate Search Problem (CSP). Specifically, our scheme uses ciphertext to participate in the operations for both machine training and data classification, which is capable of realizing homomorphic comparison without decryption, thus ensuring data privacy. The analysis illustrated that our method is provided with good portability in machine learning systems thanks to its efficiency and complete homomorphic operations.