Clustering and Aggregation for Wireless Federated Learning Under Label-Flipping Attacks

Federated Learning (FL) enables decentralized model training across distributed clients while preserving data privacy. However, its effectiveness can degrade in the presence of non-independent and identically distributed (IID) data distributions, unreliable wireless communication, and adversarial behavior such as label-flipping attacks. In this work, we propose RANGE-CFL, a robust and network-aware framework for Clustered Federated Learning that addresses these challenges. RANGE-CFL integrates a network-aware aggregation scheme to mitigate communication unreliability and introduces a dynamic clustering mechanism with participation memory to maintain group coherence under intermittent connectivity. To defend against label-flipping attacks, we develop a gradient-based client filtering method that detects and removes malicious clients using PCA-assisted clustering over output neuron gradients. Experimental results on the EMNIST and CIFAR10 datasets show that RANGE-CFL enhances global model accuracy while simultaneously reducing inter-user accuracy variance. This demonstrates its robustness against label-flipping attacks and its effectiveness in promoting fairness and stability across heterogeneous user data distributions, outperforming the baseline FL algorithm.