A Privacy-Preserving Federated Learning Scheme Using Threshold Multi-Key Homomorphic Encryption

As an emerging technology of artificial intelligence, Federated learning (FL) enables distributed data users to collectively train a global model while ensuring that the data remains local. FL has brought a new method to data-sensitive applications that require data fusion to a large extent. However, The problem of data leakage and privacy threats still exists. Multi-key homomorphic encryption (MKHE) can effectively facilitate the secure utilization of confidential data from multiple participants. In MKHE, the property of threshold allows for partial users to drop out during decryption, called failure-robust. Previous privacy-preserving federated learning (PPFL) based on MKHE schemes have seldom considered the failure-robust problem. To address this issue, we propose a threshold MK-CKKS (tMK-CKKS) scheme to design a PPFL scheme. In our scheme, a public key is generated by a joint secret key. The joint secret key is divided into secret shares using {0,1}-LSSS, and these secret shares serve as individual secret keys for each user. The model updates are encrypted with the public key before being uploaded to the cloud server for aggregation. For decryption, a subset of participants involved in the process only needs to satisfy the access structure to complete it. Our scheme can prevent privacy leakage from shared information of users and achieves threshold decryption while ensuring robustness against honest-but-curious users and collusion attacks between users and the cloud server.