Adaptive Anomaly Detection for Secure Federated Learning in IIoT Environments

Federated Learning (FL) systems train models collaboratively among Industrial Internet of Things (IIoT) devices but face prominent security issues, especially backdoor attacks due to diversified data. The presented work proposes an adaptive anomaly detection approach for security enhancement of FL in IIoT environments, considering model performance across diverse data contexts. The proposed approach leverages state-of-the-art graph-based data clustering together with differential privacy to classify data effectively into representative clusters, allowing it to detect anomalies in a localized fashion. It encompasses the most important components: hierarchical clustering, the adaptive anomaly identification module, and secure data aggregation protocol. These inventions reduce the possibility of backdoor attacks, enhance model robustness, and improve model accuracy. The role of this method was extensively validated with benchmark datasets such as Fashion-MNIST, Google Speech Commands, IoT-33, and Smart Manufacturing Systems SMS-IIOT in homogeneous and heterogeneous data settings. The results presented large reductions in backdoor attack success rates, firmly confirming the capability of the proposed solution for enhancing FL security and reliability in IIoT applications.