Aimed at the problems that computational power is limit, storage space is small, and real-time is high for requirements of drones, a method of detecting interpretable drone network intrusion based on Kolmogorov-Arnold Networks (KAN), called KIDS, is proposed. Under the inspiration of Kolmogorov Arnold Representation Theorem, KAN is to utilize spline-parameterized univariate functions for replacing the traditional linear weights to dynamically learn activation patterns, enabling effective handling of feature extraction, and achieving excellent drone network intrusion detection performance with a more lightweight network structure. Furthermore, the visualization of parameterized spline functions provides insights into the model’s decision-making process during traffic feature extraction, enhancing trust in the model’s application. The extensive experiments being over on the real-world drone network traffic dataset drone-IDS 2020, the results demonstrate that the KIDS achieves superior detection performance by still lower model complexity, and exhibits obvious generalization capability in intrusion detection for surpassing type of drones.