Abstract:Aiming at the issues of data imbalance and forgetting imbalance during training, which makes incremental intrusion detection system prone to missed detection of the specific attack classes with high-cost misclassification, an incremental intrusion detection mechanism considering cost-sensitive (IIDM-CCS) is proposed. First, an improved generative adversarial network that incorporates a detection model and an information entropy loss function for target augmentation is proposed to suppress the generation of non-specific classes and increase the proportion of specific classes in the generated samples. Second, a multi-teacher distillation method is proposed, which employs the base and old detection models to constrain the incremental training process, thereby guiding the new detection model to detect the new specific classes and alleviate the catastrophic forgetting of old specific classes. Finally, a cost-sensitive-based gradient weighted loss is designed to guide the model focus on specific classes, and a soft-label incremental distillation loss is leveraged to enhance the generalization capability on old specific classes. The comparative experimental results with other incremental intrusion detection methods show that the average recall rate of the proposed mechanism in detecting specific classes is 51.45% and 60.32% higher than those of the compared methods after incremental training. The proposed mechanism can be adapted to existing incremental intrusion detection systems to introduce cost-sensitive, thereby improving the detection accuracy for specific attack classes and mitigating the catastrophic forgetting during the incremental training.