STE-PFL: Spatial-Temporal Enhanced Personalized Federated Learning for IoT Cross-Domain Access Decision-Making

Insufficient access control in the domain of cross-domain IoT (CD-IoT) can result in unauthorized access events and impede the seamless functioning of collaborative IoT applications. The introduction of federated learning facilitates the development of dynamic intelligent access decision-making models for CD-IoT systems. Nonetheless, existing research overlooks the necessity of protecting the privacy of devices and inadequately tackles local model performance and data heterogeneity issues on CD-IoT devices. In this paper, we present a privacy-preserving access decision-making model for CD-IoT based on Spatial-Temporal Enhanced Personalized Federated Learning (STE-PFL) to address the aforementioned issues. By leveraging the Dirichlet distribution, we construct two Non-IID data scenarios using the CICIDS2018 and UNSW-NB15 datasets, respectively, to thoroughly evaluate the effectiveness of the proposed access decision-making model. The results demonstrate that our model delivers commendable accuracy in CD-IoT access decision support, and surpasses other baseline models in terms of accuracy, recall, and F1-score.