A Robustness Evaluation Method for SUDO-Based CBTC Intrusion Detection Systems

To evaluate the robustness problem of Communication Based Train Control System in intrusion detection system constructed using deep learning model, a robustness evaluation method based on symmetric uncertainty distance optimization is proposed. First, the data sample imbalance problem is solved by a hybrid sampling technique combining SMOTE and ENN, which expands a small amount of attack data and reduces the percentage of normal traffic data. Second, symmetric uncertainty is used to select the features with the highest relevance and lowest redundancy to form the optimal attack feature set. Finally, adversarial samples are generated for the feature set using distance optimization based adversarial attack. The experimental results demonstrate that the proposed SUDO adversarial attack on CNN improves the attack success rate(ASR)by 2.26%.The CNN and LSTM models used in the experiments have attack success rate of 35.34%and 23.69%.In ResNet18 model attack success rate is 31.24%and ResNet34 model attack success rate is 36.02%.