On Evaluating Black-Box Explainable AI Methods for Enhancing Anomaly Detection in Autonomous Driving Systems.

Area of Science:

• Cybersecurity
• Artificial Intelligence
• Autonomous Systems

Background:

• Autonomous vehicles (AVs) face cybersecurity risks targeting their networks.
• AI models are essential for detecting network anomalies in AVs.
• Explainable AI (XAI) is critical for understanding AI anomaly detection decisions.

Purpose of the Study:

• To introduce a comprehensive framework for assessing black-box XAI techniques in AV anomaly detection.
• To evaluate the effectiveness of global and local XAI methods in explaining AI behavior.
• To provide insights into the strengths and limitations of XAI for AV cybersecurity.

Main Methods:

• Developed a framework evaluating XAI techniques using six metrics: descriptive accuracy, sparsity, stability, efficiency, robustness, and completeness.
• Assessed two black-box XAI techniques: SHAP (SHapley Additive exPlanations) and LIME (Local Interpretable Model-agnostic Explanations).
• Applied XAI techniques to identify key features for anomaly classification on VeReMi and Sensor datasets.

Main Results:

• Evaluated SHAP and LIME against the six metrics using two AV datasets.
• Identified primary features crucial for classifying anomalous AV behavior using XAI.
• Provided a comparative analysis of SHAP and LIME performance in the context of AV anomaly detection.

Conclusions:

• The study advances the deployment of black-box XAI for real-world anomaly detection in autonomous driving.
• Offers valuable insights into the capabilities and constraints of current XAI methods in critical AV domains.
• Facilitates more robust and transparent cybersecurity solutions for autonomous vehicle networks.