I Choose You: Evaluating the Impact of Feature Selection on XAI Consensus for ML-NIDS

Machine learning-based network intrusion detection systems (ML-NIDS) are increasingly enhanced with explainable AI (XAI) techniques to support transparency and trust in automated security decisions. However, recent studies have shown that different post-hoc XAI methods often yield inconsistent explanations. These variations depended on the dataset and underlying model, and were possibly caused by training the ML models on correlated features. In this work, we investigate the hypothesis that feature selection prior to model training can influence the level of consensus among XAI methods. Through a comprehensive evaluation across multiple datasets, we analyze the impact of different feature selection strategies on explanation agreement. While we found that feature selection can improve XAI consistency in controlled synthetic settings, its effects on real-world NIDS data are mixed: occasionally enhancing, but sometimes reducing consensus, while offering only modest gains over using all features. These insights highlight the importance of thoughtful feature selection to improve interpretability and consistency in XAI-driven network intrusion detection systems.