Title: XAI-SOH-FL: Improving SOH-FL via Adaptive Aggregation and Explainable AI for Intrusion Detection in Heterogeneous IoT

Abstract:

The deployment of Intrusion Detection Systems (IDS) within Internet of Things (IoT) networks is hindered by several critical issues, including data heterogeneity, a scarcity of labeled data, and insufficient model interpretability. While Federated Learning (FL) presents a privacy-preserving alternative, current methods like SOH-FL are constrained by two major drawbacks: the necessity for manual tuning of the aggregation parameter {\gamma} and the absence of explainability in their predictive outcomes. To address these gaps, this study introduces XAI-SOH-FL, a refined framework that embeds explainable artificial intelligence and adaptive aggregation into the SOH-FL architecture.

The proposed method implements three key enhancements. First, it utilizes a dynamic {\gamma} selection process driven by similarity thresholding, allowing the aggregation mechanism to adjust to shifting data distributions. Second, it applies Bayesian Optimization to automatically identify the optimal {\gamma} values, thereby removing the burden of manual configuration. Third, it integrates SHAP (SHapley Additive exPlanations) to offer feature-level transparency regarding intrusion detection decisions.

Evaluated on the CICIDS2017 dataset, the experimental results show that XAI-SOH-FL attains an accuracy of 94.12% and an F1-score of 0.92. This performance surpasses the baseline SOH-FL model while requiring fewer communication rounds to converge. Additionally, SHAP-based analysis highlights that flow-level attributes, specifically Flow Duration and Packet Length, play a pivotal role in shaping the model’s predictions. These findings suggest that XAI-SOH-FL successfully balances accuracy, adaptability, and interpretability within heterogeneous IoT settings.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC