Spatio‐Temporal Modeling and Federated Learning‐Driven IoT Anomaly Detection: A Privacy‐Preserving Architecture for 6G Networks

Abstract

Despite 6G networks quickly maturing, the Internet of Things (IoT) will be better connected than ever before with increased capacity for big devices, and the capability to run intelligent applications in real time. Cyber incursions, data breaches, and antagonistic interferences are all made possible by the significant security holes left by these technical advancements. Because these systems face scalability limitations, privacy concerns, and dynamic IoT environmental obstacles, centralized anomaly detection technologies do not solve security risks against IoT networks. This study presents a system for real‐time threat detection using Federated Learning (FL) that safeguards data privacy and tackles these current issues. Each client separately learns data patterns in the local model training layer using a hybrid architecture that combines adversarial training to increase robustness with Graph Neural Networks (GNNs) for spatial relationships and attention‐enhanced Long Short‐Term Memory (LSTM) for temporal dependencies. After that, the locally trained models are sent to the secure federated learning layer (without sharing raw data), where a central aggregator implements security measures like consistency screening and attack mitigation in addition to performing adaptive weighted aggregation. To protect the system from evasion attempts, the defense mechanism employs an adversarial defense protocol. The proposed framework anomaly detection accuracy, F1‐score, area under the curve (AUC), and detection rate are all 97.6%. 97.2%, 98.5%, and 14.1 ms, respectively. This research shows that 6G networks may benefit from IoT systems that are safe, scalable, and privacy‐preserving when FL and GNNs operate together.