Training Policy for Privacy-Preserving Logistic Regression in Federated Learning Environments
Jorge M. Cortés-Mendoza, Andrei Nikolaevitch Tchernykh, Horacio González–Vélez · 2024
Logistic Regression (LR) is a widely used statistical model for classification problems. However, its training and evaluation in a shared environment increase the possibility of information leaking. A federated LR reduces security issues by using only locally available data for training. In a Federated Learning (FL) environment, LR receives the coefficients of local models to create the federated LR model, which is then distributed to update the local models. The exchange process does not leak confidential information when LR coefficients are encrypted. Homomorphic Encryption (HE) allows the merging of local LR models with privacy preservation (HE-LR). This work presents a novel training policy to reduce the training time with only slightly decreased quality in an FL environment with HE. We analyze the accuracy and time of FL policies with HE-LR that progressively reduce the amount of training data and exchange the LR coefficients in a privacy-preserving manner. The results show that the proposed policy can speed up the training time between 12% and 69%, compared to the traditional FL approach, with an average decrease in accuracy of 1.79% and 1.95%.