Privacy-preserving and fairness-aware secure aggregation in federated learning

Federated Learning (FL) faces a fundamental tension: cor recting demographic bias requires collecting statistics, while protect ing user privacy forbids exposing them. A recent solution resolves this through Secure Multi-Party Computation (MPC) executed on external servers, which adds deployment complexity and cost beyond what a stan dard FL system carries. We introduce SaFer-FL, a privacy-preserving federated learning framework that achieves group fairness using secure aggregation layer that protects both model updates and per-subgroup counts, from which the server derives label-aware reweighting coefficients without ever observing any individual client’s contribution. To addition ally guarantee Differential Privacy (ϵ-DP) against gradient-leakage and membership-inference attacks, each client injects a fractional share of the Symmetric Geometric Discrete Laplace (SGDL) mechanism whose per client contributions sum to a single calibrated ϵ-DP perturbation at the aggregate. Evaluated on the COMPAS and MovieLens-1M datasets with gender as the protected attribute, SaFer-FL reduces Equalized Odds and Equal Opportunity gaps relative to standard FedAvg while preserving competitive accuracy, with a controllable privacy-fairness-utility trade off.