XAI and Statistical Analysis for Reliable Intrusion Detection in the UAVIDS-2025 Dataset: From Tree to Hybrid and Tabular DNN Ensembles

Explainability is critical for security systems: The paper demonstrates that understanding why an AI model makes decisions—through techniques like SHAP (Shapley Additive explanations)—is essential for detecting intrusions in unmanned aerial vehicles (UAVs), not just achieving high accuracy scores.

Some cyberattacks are inherently difficult to distinguish: The research reveals that certain attacks like Wormhole and Blackhole attacks create network traffic patterns that overlap significantly with normal behavior, making them genuinely challenging for AI systems to detect reliably—a limitation that statistical analysis can help identify and quantify.

Rigorous statistical testing prevents false confidence: By using advanced statistical methods (Westfall-Young permutation tests and kernel density estimation), the researchers uncovered the true causes of misclassifications rather than accepting a model's performance at face value, highlighting the importance of thorough validation in security-critical applications.

Hybrid AI approaches outperform single methods: The study tested multiple machine learning approaches (tree-based models, neural networks, and combinations) and found that ensemble methods combining different techniques provided more reliable intrusion detection than any single approach alone.

Transparency enables trust in automated security: By analyzing which features (network characteristics) each attack type targets and visualizing how the model makes decisions, the research provides security experts with actionable insights to understand and potentially improve UAV defense systems.