FedRec-RDP-HE with HMAC: A Privacy-Preserving Federated Recommendation Framework Using RDP and Additive Homomorphic Encryption
Year: 2026
Type: conference-paper
Abstract: This paper presents FedRec-RDP-HE, a privacy-preserving federated recommendation framework built on top of a matrix factorization backbone and reinforced with three complementary protection layers: Rényi Differential Privacy (RDP), elliptic-curve-based additive Homomorphic Encryption (HE), and HMAC-based integrity verification. In the proposed architecture, raw user ratings never leave the client side. Instead, each selected client performs local matrix factorization training on private interactions, computes item-update deltas, applies ℓ 2 clipping, and injects Gaussian noise calibrated according to an RDP-based mechanism before transmission. To further protect communication, the privatized updates are encrypted on the client side through an ECC-based additive homomorphic encryption layer, while an HMAC tag is generated to guarantee message integrity during transmission. On the server side, received messages are first verified, encrypted coordinates are aggregated, aggregated ciphertexts are decrypted, and the resulting updates are used to refine the global item matrix. Experimental results on MovieLens 100K and MovieLens 1M show that the proposed RDP-based mechanism achieves lower MAE and RMSE than a Laplace-based LDP counterpart under the same federated matrix factorization workflow. These results highlight the benefit of Gaussian perturbation for improving the privacy–utility trade-off within the proposed FedRec-RDP-HE framework.
Source: Procedia Computer Science
Institution: Tunis University
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: diamond
OpenAlex ID: W7225110259