Personalized Federated Learning With Multigranularity Confidence Alignment for IoT Device Collaboration

Federated learning (FL) provides a privacy-preserving solution for model training across distributed Internet of Things (IoT) devices. IoT scenarios typically involve highly heterogeneous data, limited computational capacity, and unstable communication links. Non-independent and identically distributed (non-IID) data further exacerbate training difficulties. Common challenges include reduced model generalization, insufficient robustness against abnormal inputs, and performance unfairness among clients. To address these challenges, personalized federated learning with multi-granularity confidence alignment (FedMGCA) is introduced. FedMGCA incorporates a multi-granularity confidence alignment mechanism to calibrate model confidence at the feature, decision, and distribution levels. In addition, a trust-based dynamic aggregation strategy is adopted to reweight client updates based on reliability assessments. The experimental results demonstrate that FedMGCA consistently outperforms existing personalized federated learning algorithms across various benchmarks.