Evaluation of Explainable Artificial Intelligence in IoT Intrusion Detection Systems Under DeepFool Adversarial...

Area of Science:

• Cybersecurity
• Artificial Intelligence
• Machine Learning

Background:

• Increasing complexity of Internet of Things (IoT) systems necessitates transparent and trustworthy machine learning intrusion detection systems (ML-IDS).
• Post hoc explainable AI (XAI) methods, such as SHapley Additive exPlanations (SHAP) and Local Interpretable Model-agnostic Explanations (LIME), are commonly used to interpret ML models.
• The robustness of these XAI methods against adversarial conditioning remains an understudied area.

Purpose of the Study:

• To propose a unified system for evaluating explanation fidelity using sparsity, completeness, and robustness metrics.
• To benchmark the robustness of SHAP and LIME under adversarial perturbations using the DeepFool algorithm.
• To assess the impact of adversarial attacks on the fidelity of explanations provided by SHAP and LIME.

Main Methods:

• Evaluation of explanation fidelity using three metrics: sparsity, completeness, and robustness.
• Application of minimally distorting DeepFool input perturbations to assess robustness.
• Benchmarking SHAP and LIME across three datasets (BoT-IoT, Edge-IIoT, N-BaIoT) and four classifiers (CNN, DNN, LSTM, RF).

Main Results:

• A consistent trade-off was observed between SHAP and LIME under attack.
• SHAP demonstrated stronger feature alignment and higher completeness against adversarial perturbations.
• LIME showed greater rank stability in top-k feature overlap but produced more spurious attributions and offered less explanatory power than SHAP, especially with synthetic features.

Conclusions:

• High model accuracy does not guarantee high explanation fidelity, particularly under adversarial conditions.
• Robustness-aware XAI is essential for reliable cybersecurity applications.
• The study provides reproducible parameters to guide the adoption of XAI in adversarial environments.