Trust-Aware Reinforcement Selection for Robust Federated Learning under Adaptive Adversaries

Federated learning (FL) has emerged as a promising framework for privacy-preserving collaborative training, yet the presence of Byzantine clients poses a critical challenge for robust aggregation. Existing defenses such as FedAvg, Krum, Trimmed Mean/Median, FLTrust, and SARA exhibit significant performance drops in dynamic or mixed-attack environments. In this paper, we propose TARS, a Trust-Aware Reinforcement Selector for robust FL aggregation under adversarial and non-IID conditions. TARS leverages a trust-regularized Q learning strategy to dynamically select the optimal aggregation rule in each round, accounting for model performance and trustworthiness signals. Experimental results on MNIST and CIFAR-10 with 20% Byzantine clients demonstrate that TARS consistently outperforms all baselines, achieving a final test accuracy of 97.7% on MNIST and 80.5% on CIFAR-10, surpassing FLTrust and SARA by at least 2.5% and 3.6%, respectively, as shown in Table IV. TARS also achieves the highest optimal rule selection rate (93.6% on MNIST, 89.8% on CIFAR-10), robust convergence, and resilience against both label flipping and Gaussian attacks. These results establish TARS as a mathematically principled and empirically validated solution for trustworthy federated learning in adversarial settings.