Dual-Server Privacy-Preserving Collaborative Deep Learning: A Round-Efficient, Dynamic and Lossless Approach

To address limitations in existing privacy-preserving collaborative deep learning (CDL) schemes, we propose a dualserver privacy-preserving CDL scheme based on homomorphic encryption and a masking technique. Specifically, in our scheme a random seed is used to initialize a pseudorandom generator that produces multiple pseudorandom numbers. These pseudorandom numbers, along with a random noise, are utilized to generate masks that are added to all parameters of a participant’s locally trained model. By using homomorphic encryption, the random noise can be encrypted and eventually used to remove the masks with low message expansion. This also ensures that the global model is lossless in accuracy. Furthermore, if participants join or leave the system, only the time required to complete both model update aggregation and encrypted masks aggregation is affected. We demonstrate that our scheme is round-efficient, dynamic and lossless. We also show that it is secure against inference attacks and can resist collusion attacks of up to t − 2 participants and one of the two servers, where t is a security parameter indicating the minimum number of participants that participate in an aggregation round.