A Verifiable Privacy-Preserving Federated Learning Algorithm Based on Functional Encryption
Jiayi GU , Ke ZHANG , Xizhao LUO , Chongben TAO , Jian WENG
Federated learning faces two major security challenges: privacy leakage from shared model updates and malicious submissions that compromise aggregation correctness. To address these challenges, this paper proposes DGAg-GV, a privacy-preserving and verifiable distributed federated learning framework. First, based on an existing verifiable decentralized MCFE construction, DGAg-GV instantiates concrete verification procedures for ciphertexts and functional key shares in federated learning, enabling the detection of invalid ciphertexts, forged functional key shares, and incorrect round labels while preserving input privacy. Second, it introduces label binding and dynamic grouping with temporary identities to resist cross-round mix-and-match attacks and multi-round privacy leakage. Third, it proposes dynamic grouping and group-level consistency checking to reduce DK-share verification overhead and improve large-scale verification efficiency through group-level verification and efficient rejection of invalid group submissions. Experiments on MNIST and CIFAR-10 show that DGAg-GV effectively detects invalid client submissions and excludes affected groups while maintaining model accuracy comparable to baseline schemes. Complete-workflow experiments show that group-level consistency checking reduces DK-share verification time by approximately 65.5%-67.5%, while protocol-level stress tests from 100 to 1000 clients demonstrate approximately linear estimated verification overhead. These results indicate that DGAg-GV provides a feasible approach toward privacy-preserving and verifiable large-scale federated learning.
federated learning / privacy protection / verifiable computation / functional encryption / malicious clients
Higher Education Press 2026
/
| 〈 |
|
〉 |
PDF
(1491KB)
