Anti-Collusion and Latency-Tolerant Privacy-Preserving Federated Learning Framework via Multi-Key Homomorphic Aggregation

In this paper, we propose MKHA-PPFL, a federated learning framework that utilizes advanced multi-key homomorphic aggregation (MKHA) to enhance security guarantees in a cryptographic manner. MKHA enables multiple participants to encrypt their data with individual keys while allowing homomorphic operations to be performed on the combined ciphertexts without requiring a shared key. In MKHA-PPFL construction, we invite an authorized center to manage the participants' secretpublic key pairs. The proposed model achieves the following objectives: (1) each participant's weight information is completely shielded from the server and other participants; (2) server-side leakage is fully avoided; (3) the global weight parameter can be updated timely by the server, and (4) resistance to quantum computing attacks is guaranteed by incorporating post-quantum cryptographic techniques. These advantages significantly enhance the overall system security by an efficiency/privacy tradeoff.