Noise-Robust Federated Learning via Contrastive Initialization and Sparse Regularization

Federated learning (FL) enables collaborative model training across decentralized clients without sharing raw data, making it highly suitable for privacy-sensitive applications. However, the presence of noisy labels, particularly under nonindependent and identically distributed (non-IID) data partitions, poses a significant challenge, often leading to biased local updates and poor global generalization. In this work, we propose Sparse Regularization-based Robust Federated Learning with Contrastive Pre-training (FedSRCP), a unified framework specifically designed to address label noise in decentralized, non-IID FL environments. FedSRCP integrates server-side contrastive pretraining for label-free representation learning with client-side Generalized Cross-Entropy (GCE) loss and sparse regularization, collectively enhancing robustness to corrupted supervision while preserving model generality. A detailed ablation study reveals the complementary roles of these components in mitigating overfitting and promoting discriminative feature learning. Extensive experiments on benchmark datasets including SVHN, CIFAR10, CIFAR-100, CIFAR-100N, and Clothing 1 M demonstrate that FedSRCP consistently outperforms standard FL algorithms and noise-robust baselines under both synthetic and real-world noise conditions. Notably, FedSRCP achieves up to $5 \%$ accuracy improvements in high-noise regimes while maintaining strong performance on clean data. These results highlight FedSRCP’s practical value for robust federated learning in large-scale, noisy, and heterogeneous environments.