MO-Transformer-IDS: Multi-objective Optimization-based Transformer for Multi-Classification Intrusion Detection of Cyber-Physical Power Systems

Deep neural networks have been widely applied in intrusion detection systems (IDSs), but most of the existing models are designed manually and it is difficult to obtain a lightweight and high-performance model, especially for multi-classification IDS. This work presents an automatic deep learning-based IDS method, i.e., a multi-objective optimization-based transformer for multi-classification intrusion detection of cyber-physical power systems (CPPS), termed MO-Transformer-IDS. This method considers both the architectures such as the number of attention heads in the Transformer-Encoder, the type of activation function, and the number of layers in the Encoder, and the hyperparameters such as the learning rate, batch size, and dropout rate as the decision variables. The two optimization objectives are to minimize the detection error rate and the number of transformer model parameters. The evolutionary operations under the framework of multi-objective differential evolution (MODE) are designed elaborately to search for a Pareto-optimal transformer model. The experimental results on the thirty-seven-classification IDS issue of Power System Attack Datasets have demonstrated that the proposed MO-Transformer-IDS outperforms three state-of-the-art automatic deep learning (ADL) methods including SOPA-GA-CNN, WOA-ANN, GA-CNN, and an advanced LSTM in terms of Accuracy, Precision, Recall, F1-Score, and the number of transformer model parameters.