Energy-Aware Hierarchical Kolmogorov–Arnold Networks for TinyML-Based UAV Intrusion Detection on Microcontrollers

The increasing use of Unmanned Aerial Vehicle (UAV) networks in mission-critical applications requires intrusion detection systems (IDS) that maintain reliable security functionality under strict energy, memory, and real-time constraints. While hierarchical intrusion detection has been explored, existing approaches largely ignore the impact of inference energy consumption on battery-powered UAV platforms, where excessive computation directly reduces mission duration. This paper presents H-MicroKAN, an energy-aware hierarchical intrusion detection architecture based on Kolmogorov-Arnold Networks (KAN), specifically designed for deployment on microcontroller-class UAV hardware. The proposed system employs a two-stage hierarchical inference strategy that exploits the empirical observation that 70-90% of operational UAV traffic is benign. Stage 1 uses an ultra-lightweight binary classifier (926 parameters) to detect malicious activity, while Stage 2 activates a more expressive multi-class attack classifier (5,126 parameters) only when an attack is detected. Unlike prior TinyML-based IDS that apply uniform inference costs to all traffic, H-MicroKAN enables traffic-aware conditional execution, significantly reducing unnecessary computation while preserving detection reliability. Evaluation on the UAVIDS-2025 dataset shows that the system achieves 99.0% attack detection accuracy and 94.47% attack classification accuracy with only 6,052 parameters (23.6 KB). Hardware measurements on an ESP32 validate real-time operation and demonstrate up to 81% energy savings on benign traffic. These results indicate that hierarchical intrusion detection with conditional inference provides a practical security mechanism for energy-constrained UAV systems operating under adversarial conditions.