Efficient and Privacy-Preserving Federated Learning with Gradient Control against Data Poisoning

Abstract

Federated Learning (FL) enables collaborative model training across decentralized clients while preserving data privacy. However, practical deployments are often limited by significant communication overhead and vulnerability to Byzantine poisoning attacks. Existing defenses typically rely on post-hoc anomaly detection, but executing complex distance metrics or clustering on encrypted, sparsified updates creates a substantial computational burden for the aggregation server. We present a privacy-preserving FL framework that addresses these challenges by integrating Top-ksparsification, non-interactive zero-knowledge proofs (NIZKPs), and homomorphic encryption. Instead of relying on expensive ciphertext distance computations, our architecture uses a pre-aggregation global mask sign vector, generated through majority voting, to filter anomalous updates. This mechanism treats unselected gradient coordinates as explicit zero-votes, which mitigates malicious coalitions attempting to manipulate disjoint parameter subsets. A local error feedback mechanism ensures that heterogeneous client updates align over successive training rounds. Combined with NIZKPs to enforce coordinate-wise magnitude bounds, the framework provides Byzantine resilience without increasing communication costs or compromising privacy. Evaluations on MNIST and CIFAR-10 show that our approach maintains high communication efficiency and robustness. Under a 40% malicious client poisoning attack and a 50% sparsification ratio, the framework achieves final accuracies of 92.14% and 63.20%, respectively, demonstrating its effectiveness in bandwidth-constrained, hostile environments.