A temporal federated learning of neural network for enhancing controllability temporal network robustness on federated social internet of things networks

This paper presents a novel Temporal Federated Learning framework to enhance the controllability robustness of Federated Social Internet of Things (FSIoT) networks under dynamic and adversarial conditions. The proposed model, termed Recovery Controllability using Temporal Federated Learning (RCTFL), combines Long Short-Term Memory (LSTM), temporal embeddings, and federated aggregation to predict and reconstruct failing links without sharing raw data. The RCTFL ensures privacy preservation, scalability, and improved resilience against targeted and random attacks by distributing the learning process across multiple IoT domains. Using the SmartSantander smart-city datasets, the framework was evaluated under various attack scenarios, including targeted node removal, link disruption, and cascading failures. Experimental results demonstrate that RCTFL outperforms conventional method baselines by up to 10% in both accuracy and F1-score, while requiring fewer structural modifications and achieving 45% reductions in computation time. The findings confirm that integrating temporal federated learning with controllability theory enables distributed IoT systems to maintain stable operation and recover efficiently after disruptions.