Privacy-preserving two-parties logistic regression on vertically partitioned data using asynchronous gradient sharing

The full application of machine learning has caused plenty of problems with privacy-preserving. Especially in multi-party machine learning, private data is often exposed in the aggregation,transmission, and communication phase, which leads to the problem of private data leakage. Existing works use secure multi-party computing (SMPC) or secret-sharing technology to ensure the privacy-preserving of multi-party machine learning. Nevertheless, it brings enormous cost and feasibility drawbacks. The partition method of datasets is one of the most critical factors affecting the performance of machine learning. Vertically partitioned data has the problems of incomplete feature information held by a single participant and complicated training process. Therefore, it has to be tackled urgently that how to efficiently and safely complete the multi-party training using vertically partitioned datasets. Moreover, training logistic regression models efficiently is one of the directions worth working on. In this paper, we propose a protocol using that can complete the logistic regression modeling of vertically partitioned data by asynchronous gradient sharing. At the same time, we use an efficient homomorphic encryption method to protect private data. The experiments show that our protocol can reduce the training time in the case of a small impact on the output results, and speedup can be over 10x. Meanwhile, it will ensure the security of the vertically partitioned dataset.