Privacy-Preserving Logistic Regression Model Training Scheme by Homomorphic Encryption

Privacy-Preserving Logistic Regression Model Training Scheme by Homomorphic Encryption

연구 분야: Cryptography

학회: International Conference on Information and Communications Security

In the field of big data, logistic regression for binary and multi-class classification is widely used. Nowadays, there is growing concern about data privacy protection issues. This paper focuses on scenarios involving two parties participating and data being horizontally distributed. Based on homomorphic encryption, a logistic regression model training scheme is designed. This scheme reduces the number of iterations in the training process by using the second-order approximation Newton’s method. It employs the conjugate gradient method to solve the Newton’s method, and introduces a small amount of interaction to reduce ciphertext domain division operations, greatly reducing the computational overhead of the ciphertext domain. Additionally, a new encoding method is used to reduce the number of ciphertext multiplications and communication overhead. Furthermore, the “One-vs-Rest" decomposition strategy is adopted, combined with SIMD(Single Instruction, Multiple Data) technology, extending the binary model to multi-classification. Experimental results show that with the use of Privacy-preserving Logistic Regression scheme (PPLR), for most datasets, setting the number of iterations to within 3 rounds can achieve accuracy comparable to existing privacy protection schemes of 5 to 7 rounds. For sample datasets with 60 and 112 dimensions, existing similar schemes require 90 s and 165 s to complete 5 rounds of iteration, while this scheme only requires 8 s and 27 s. Moreover, the communication overhead is reduced to half of the original scheme, requiring only 30.8MB and 62.7MB to complete training.