Secure Logistic Regression Based on Homomorphic Encryption: Design and Evaluation

• Published in: JMIR medical informatics Vol. 6; no. 2; p. e19
• Main Authors: Kim, Miran, Song, Yongsoo, Wang, Shuang, Xia, Yuhou, Jiang, Xiaoqian
• Format: Journal Article
• Language: English
• Published: Canada JMIR Publications 17.04.2018
• Subjects: Approximation; Biomedical research; Efficiency; Machine learning; Methods; Neural networks; Original Paper; Outsourcing; Privacy; Regression analysis; homomorphic encryption; logistic regression; machine learning; gradient descent
• ISSN: 2291-9694; 2291-9694
• DOI: 10.2196/medinform.8805

Learning a model without accessing raw data has been an intriguing idea to security and machine learning researchers for years. In an ideal setting, we want to encrypt sensitive data to store them on a commercial cloud and run certain analyses without ever decrypting the data to preserve privacy. Homomorphic encryption technique is a promising candidate for secure data outsourcing, but it is a very challenging task to support real-world machine learning tasks. Existing frameworks can only handle simplified cases with low-degree polynomials such as linear means classifier and linear discriminative analysis. The goal of this study is to provide a practical support to the mainstream learning models (eg, logistic regression). We adapted a novel homomorphic encryption scheme optimized for real numbers computation. We devised (1) the least squares approximation of the logistic function for accuracy and efficiency (ie, reduce computation cost) and (2) new packing and parallelization techniques. Using real-world datasets, we evaluated the performance of our model and demonstrated its feasibility in speed and memory consumption. For example, it took approximately 116 minutes to obtain the training model from the homomorphically encrypted Edinburgh dataset. In addition, it gives fairly accurate predictions on the testing dataset. We present the first homomorphically encrypted logistic regression outsourcing model based on the critical observation that the precision loss of classification models is sufficiently small so that the decision plan stays still.