An AI-Driven Framework for Intrusion Detection and Predictive Threat Modeling in UAV Ecosystem

The escalation of Internet of Things (IoT)-enabled unmanned aerial vehicle (UAV) networks and autonomous vehicles has increased their exposure to cyber threats, highlighting the need for explainable intrusion detection systems (IDSs). In this paper, we propose a framework that combines a Transformer-based model, reinforcement learning (RL)- based decision thresholding, and explainable artificial intelligence (XAI) using Local Interpretable Model-agnostic Explanations (LIME). The framework is tested on UAVIDS-2025, which includes 122,171 flows from four attack categories (Blackhole, Wormhole, Sybil, and Flooding) and Normal Traffic. A leakage-free protocol is used, with feature scaling and SMOTE-based class balancing applied only to the training partition, while the validation and test partitions are held out. The accuracy, F1-score, and ROC-AUC of the model for binary intrusion detection are 97.73%, 98.53%, and 99.97%, respectively, with a false positive rate (FPR) of 0.10% and a false negative rate (FNR) of 2.86%. The multi-classifier with a shared Transformer backbone achieves 92.72% accuracy and a macro F1 score of 92.87%. There are 399,745 trainable parameters in the architecture, and a footprint of 1.54 MB. Ablation studies and comparisons with five machine-learning baselines validate the design choices. LIME has shown it is accurate, interpretable, reliable, and suitable for use with UAVs by identifying features such as flow duration, packet/byte rates, and drop-rate statistics.