Federated Learning Privacy-preserving Scheme Based on DMK-CKKS

This paper proposes a privacy-preserving federated learning (FL) framework by intorducing an enhanced multi-key homomorphic encryption scheme based on MK-CKKS. To address the vulnerabilities of original MK-CKKS against semi-honest participants and server collusion attacks in FL, we redesign the key architecture by introducing an aggregated public key derived from all participants' private keys, and propose the improved Decentralized MK-CKKS(DMK-CKKS) algorithm. This modification eliminates single-point decryption risks while reducing communication rounds through merged aggregation and collaborative decryption phases. The proposed framework ensures computational security under the Ring Learning with Errors (RLWE) assumption and resists collusion attacks involving up to k<N-1 adversaries. Comparative experiments with Paillier and MK-CKKS baselines demonstrate significant improvements: 1) Communication costs are reduced by 37% compared to MK-CKKS (375 KB vs. 526 KB per client); 2) Computational efficiency surpasses Paillier by an order of magnitude in encryption/decryption latency; 3) Model accuracy remains comparable (94.47% vs. 95.39% baseline) despite cryptographic overhead. Theoretical security proofs and empirical evaluations validate the scheme's robustness and practicality for privacy-sensitive FL scenarios.