A Comparative Analysis of DNN-based White-Box Explainable AI Methods in Network Security

White-box explainability methods outperform black-box approaches for security applications: The study found that white-box XAI techniques (LRP, IG, and DeepLift) scored significantly higher in robustness and completeness—metrics the researchers identified as crucial for intrusion detection systems, suggesting these methods are more reliable for security-critical decisions.

Explainability is essential for security professionals to trust AI: The research emphasizes that security analysts need to understand why AI systems flag potential intrusions rather than simply accepting automated decisions, highlighting that explainability isn't just a nice-to-have feature but a requirement for responsible deployment of AI in cybersecurity.

Multiple evaluation metrics matter for responsible AI assessment: Rather than relying on a single performance measure, the study evaluated six distinct metrics (descriptive accuracy, sparsity, stability, robustness, efficiency, and completeness), demonstrating that comprehensive evaluation across multiple dimensions is necessary to properly assess whether an AI system is trustworthy and suitable for real-world use.

Standardized frameworks improve AI accountability: By developing and open-sourcing an end-to-end evaluation framework tested across three different datasets, the researchers created tools that allow the broader community to independently verify and improve XAI methods, supporting transparency and reproducibility in responsible AI development.