Practical Federated Learning for Samples with Different IDs

Abstract

Federated learning is widely used in various fields, it usually consists of sample alignment phase and training phase, where sample alignment is the first phase. For example, in horizontal federated learning, if the databases of parties contain some identical samples, then parties could use sample alignment to remove these duplicate samples before training, and in vertical federated learning, parties are required to use sample alignment to put the samples of the same user on the same row of both databases before training. Meanwhile, the current sample alignment schemes in federated learning are almost ID-based, and they assume the two participants have the same ID. Consider that these schemes cannot deal with the sample alignment problem for samples with different IDs, we present a sample alignment scheme that allows two participants with different IDs to align their samples. Our sample alignment scheme is based on Oblivious Programmable PRF (OPPRF), which doesn’t have much public key operation. After aligning the samples utilizing our scheme, the two participants could accomplish a variety of secure two-party machine learning tasks. In this paper, we design the privacy-preserving logistic regression training scheme using additive homomorphic encryption, thus achieving the whole federated logistic regression process. We implement our sample alignment scheme to verify the efficiency, and the experiments show that our sample alignment scheme only requires 216 s when the set sizes of sender and receiver are 224 and 220. Besides, we conduct experiments to verify the feasibility of our logistic regression training scheme.