XAI-based Feature Selection for Improved Network Intrusion Detection Systems

Explainability gaps exist in current security systems: The paper identifies that explainability and evaluation of AI models are "lacking" in intrusion detection systems, meaning security professionals often cannot understand why AI systems flag certain network activity as threats—a significant gap for responsible deployment of AI in critical infrastructure.

XAI methods improve both accuracy and transparency: The research demonstrates that AI models "perform better" when using explainability-based feature selection compared to traditional methods, suggesting that making AI systems more interpretable doesn't sacrifice performance and may actually enhance it.

Understanding "why" helps security experts do their jobs: By using XAI techniques to identify which network features matter most for detecting attacks, the work helps "security analysts" better understand "critical intrusion traits" and the reasoning behind AI decisions, enabling human experts to validate and trust the system's recommendations.

Feature selection is foundational to responsible AI in security: The paper emphasizes that identifying the most important data features is "essential" for both explainability and evaluation, suggesting that responsible AI in security starts with transparency about what information the system actually uses to make decisions.

Open-source frameworks enable community accountability: By making source code publicly available, the authors support broader scrutiny and development of XAI-based security tools, allowing the security community to verify methods and build upon transparent foundations rather than relying on "black box" systems.