Robust and Privacy-Preserving Federated Learning Against Malicious Clients: A Bulyan-Based Adaptive Differential Privacy Framework

Federated learning (FL) enables collaborative model training across decentralized clients without sharing raw data. However, malicious or byzantine clients may compromise global model integrity through adversarial updates, while the addition of differential privacy (DP) noise can significantly reduce accuracy if not carefully managed. To address these dual challenges, we propose a robust and privacy-preserving FL framework featuring three core innovations: (i) a Bulyan-based aggregator that discards outlier gradients to neutralize adversarial behaviors, (ii) a GroupNorm-based convolutional neural network (CNN) design for DP compatibility, and (iii) an adaptive noise-scheduling mechanism that gradually reduces noise variance across training rounds. Specifically, we configure the Bulyan aggregator with $k=7$ and retain $K=5$ gradients per round, using a trim ratio of $\gamma =0.2$ . The DP noise follows an exponential decay schedule $\sigma _{r} = 0.30\exp (-0.01\,r)$ . Experimental results on CIFAR-10 under a 10% malicious-client scenario demonstrate that our method consistently mitigates poisoned updates and achieves a final global accuracy of 23.58% while maintaining a rigorous privacy budget of $\varepsilon =1.5$ , $\delta =10^{-5}$ . Despite an unavoidable performance gap relative to non-private or non-adversarial baselines, legitimate clients reach local accuracies near 80%. This outcome underscores the interplay between robust aggregation and flexible DP tuning to preserve model integrity and privacy. Our framework thus paves the way for future research on advanced aggregator heuristics, per-layer noise calibration, and refined defenses against sophisticated adversaries in large-scale FL.