Search for a command to run...
The intense development of cyber threats has revealed the weaknesses of traditional intrusion detection systems (IDS) that fail to keep up with adversarial attacks that aim at utilizing the vulnerabilities of machine learning models. In an effort to overcome this problem, we suggest an Adversarial Robust Intrusion Detection framework based on Graph Neural Networks (GNNs). In contrast to traditional IDS, which view network traffic as discrete data points, our framework represents network communication as a graph with complex dependencies among entities, e.g. between hosts, connections and protocols. We present adversarial training methods that are robust to evasion and poisoning attacks, which provide robustness in non-beneficial settings. This architecture combines node- and edge-level feature extraction, relational learning with graph convolution, and anomaly-sensitive classifiers, which are type optimized using adversarial perturbations. The outcome of our experiments on standard intrusion datasets indicates that our model has better detection accuracy and resilience in relation to the latest deep learning baselines, even in a severe adversarial environment. The paper also indicates the possibilities of jointly using GNNs with adversarial defences to develop future-generation intrusion detection systems, which are both adaptive and resilient to current cyber defence infrastructures.