Abstract:With the establishment and standardization of data exchange markets,a new demand in which multiparties collaboratively train a machine learning model has emerged. Federated learning enables multiple data owners to jointly train a model,with the requirement that the model is trained and shared by all participants. The existing federated learning frameworks can not be applied to scenarios in which the data owners and model demander have different requirements and the model is jointly trained but not shared. A collaborative computing scheme for privacy-preserving logistic regression based on homomorphic encryption is proposed,which is independent of any third-party computing platforms. The collaborative compuing scheme includes a multiparty collaborative computing framework that comprises multiple data owners,a model demander,a key generator,and an interactive collaborative computing process based on the framework. With this framework,a model can be collaboratively trained without the leakage of model information or data privacy. The security of the collaborative computing scheme is analyzed by establishing an attack model. Based on the advanced floating-point fully homomorphic encryption scheme called the Cheon-Kim-Kim-Song（CKKS）,a prototype system is implemented on a small computer cluster. This is optimized for calculations and communication,including the early termination of the training process and offloading the ciphertext homomorphic operations to the Graphics Processing Unit（GPU） to improve computational efficiency. The experimental results show that the computational optimizations can improve the system performance by approximately 50 times,and the prototype system can satisfy the practical requirements for small and medium-sized data sets.

Collaborative Computing of Privacy-Preserving Logistic Regression Based on Homomorphic Encryption