Privacy-Preserving Multi-User Joint Data Logistic Regression Inference Scheme

Nowadays, in the era of big data with rich sources of information, machine learning technologies are developing rapidly, and they are widely used in areas such as medical treatment, genomic prediction, spam detection, face recognition, and financial prediction. As a classification model in machine learning, the logistic regression model has become an industrial focus due to the simplicity and efficiency of algorithms. In this context, the “model as a service” business model has emerged. Specifically, companies with expertise and large-scale training data provide high-quality models to provide reasoning services for users. However, in machine learning classification models, it is important that feature data and inference models remain private. In view of this problem, the existing privacy inference technology considers that the single user has all the feature data, while the feature data of some scenarios exist between multiple users in the form of segmentation. In this paper, based on the fully homomorphic encryption CKKS scheme and the secure computing technology, this paper designs a privacy logic regression inference scheme suitable for multi-party data combination, which can simultaneously protect the model privacy of the server side, the data feature privacy of each user and the access rights of the inference results. Finally, experimental tests of the designed scheme are performed to verify its high performance.