An adaptive graph neural network-based intrusion detection system for airborne network

With the rapid development of airborne information technology, the security of airborne networks has received more attention. To provide effective protection for airborne networks, we propose an adaptive graph neural network intrusion detection system; containing a graph generation process and a graph reconstruction process. Firstly, there is inherent uncertainty in the classification boundary for imbalanced datasets in airborne networks. We propose a Weight-based k-nearest neighbor to assign differential weights, thereby constructing the adjacent weight matrix. Then, we introduce a Gaussian graph aggregation classification method to improve the accuracy of the initial graph construction. In this process, the data with connected relationships are mapped into a low-dimensional Gaussian distribution through aggregation. We sample from the Gaussian distribution for each node data to establish new connections through the latent variables for anomaly intrusion detection. Finally, the proposed method was tested on public network intrusion detection and aviation data bus intrusion detection to verify its effectiveness. The experimental results demonstrate that the proposed method improves model classification performance by using an adaptive graph neural network that repeatedly aggregates features from similar nodes, leading to more accurate detection results.