T9-FL: An Adaptive XAI-driven Federated Learning Framework for Privacy-preserving Intrusion Detection in Non-IID Edge IoT Networks

Industrial IoT (IIoT) faces significant security challenges due to the decentralization of Industry 5.0.In this paper, T9-FL, a robust Federated Learning framework for detecting advanced cyberattacks is introduced.Our framework harnesses the capability of T9-ResNet along with FedProx optimization and Local SMOTE balancing techniques to mitigate Non-IID statistical heterogeneity and extreme class imbalance.Evaluated on the Edge-IIoTset across 15 attack categories, T9-FL achieved a global accuracy of 81.6% and a balanced accuracy of 79.58%, outperforming standard FedAvg by 4.2%.To validate generalizability, the framework was further tested on the CICIoT2023 dataset, achieving 78.60% accuracy.Ablation studies confirm that T9-ResNet's architecture is effective, as removing skip connections led to a significant 19.17% performance drop.Additionally, interpretability at featurelevel is offered by SHAP interaction values computed from network packet signatures.T9-FL offers robust, interpretable, and scalable security solutions finding the balance between rigid data privacy and communication requirements with high-tier anomaly detection.