A Guided Kolmogorov-Arnold Network for Enhanced Security in Assistive UAV Technologies

The increasing reliance on unmanned aerial vehicles (UAVs) in mission-critical assistive technology applications has elevated the need for secure and resilient communication. Intrusion Detection Systems (IDS) play a pivotal role in safe-guarding UAV networks, yet conventional neural architectures often struggle with high-dimensional telemetry, fixed activation functions, and limited explainability. In this paper, we propose the Guided Kolmogorov-Arnold Network (GS-KAN), a novel IDS framework that replaces static activations with learnable B-spline transformations and systematically optimizes architectural hyperparameters—width, spline degree, and grid resolution—via a grid-based search strategy. We introduce a unified evaluation metric to guide multi-objective performance trade-offs across accuracy, precision, recall, and F1 score. Experiments on three benchmark intrusion datasets (CIC-IDS 2017, CIC-IDS 2018, and CIC-DDoS 2019) demonstrate that GS-KAN consistently outperforms standard MLPs and KAN variants, achieving robust generalization across optimizers and data distributions essential for reliable deployment in assistive technology environments.