基于对比表示蒸馏的轻量化异常检测方法OA

To address the challenge of real-time deployment for network traffic anomaly detection models in resource-constrained environments,this paper proposes a lightweight anomaly detection method based on contrastive representation distillation.First,the teacher model integrates local and global temporal convolutions for multi-scale temporal feature extraction,and incorporates residual connections and an at-tention mechanism to enhance feature propagation,mitigate gradient vanishing,and focus on critical in-formation.Simultaneously,a hybrid architecture combining convolutional neural networks(CNN)and long short-term memory networks serves as the classifier to further integrate spatiotemporal features and improve classification performance.Second,the student model employs a streamlined CNN backbone in-tegrated with gated recurrent units to enhance its spatiotemporal feature extraction capability.Finally,within the contrastive representation distillation framework,the discriminative knowledge of the teacher model regarding the differences between positive and negative samples is effectively transferred to the stu-dent model.Experimental results demonstrate that the proposed method significantly reduces the number of model parameters,accelerates the training of the student model,and enhances the detection perfor-mance by preserving the discriminative capability of the teacher model.