LBSFL: a lightweight robust federated learning method for IoT

Federated learning (FL) provides a privacy-preserving paradigm for distributed model training in Internet-of-Things (IoT) systems, but it faces challenges from non-IID data heterogeneity and adversarial threats. Although the Shapley value provides a principled framework for client contribution evaluation, its exponential complexity renders it impractical for resource-constrained IoT devices. This paper proposes a Lightweight Bandit-Shapley Federated Learning (LBSFL), a lightweight trust-aware FL framework tailored for IoT environments. A truncated Monte Carlo method efficiently approximates client contributions under limited computational budgets. A multi-armed bandit mechanism balances exploration and exploitation during client selection, thereby reducing communication overhead. Contribution-guided aggregation weights suppress unreliable updates while prioritizing high-quality clients. Experiments under heterogeneous data distributions and adversarial scenarios show that LBSFL achieves competitive model accuracy while substantially reducing computational and communication overhead in most evaluated settings. These results indicate that LBSFL provides a favorable trade-off between robustness and efficiency for IoT-oriented federated learning.