A Network Intrusion Detection System Based on Self-supervised Co-contrastive Learning

Leveraging graph structures, Graph Neural Networks (GNNs) have demonstrated promising potential in network intrusion detection applications. However, most existing GNN-based intrusion detection systems depend on labeled network traffic, which greatly limits their practical applications, as the process of obtaining labeled data samples in a wide range of computer networks is costly and labor intensive. Furthermore, limited labeled traffic often fails to adapt to unfamiliar attack scenarios. In order to solve these challenges, we introduce CoGN, a network intrusion detection method using co-contrast mechanism, by combining pre-training of embedded learning with unsupervised detection algorithm. Our model transforms computer network traffic into a graph structure and extracts local and high-order features from the network pattern and meta-path perspectives via the graph neural network. These two views collaborate and supervise each other to learn their own characteristics to arrive at the final high-level network traffic embedding. The experimental results show that the proposed intrusion detection system can effectively mitigate most of the performed attacks.