Lightweight anomaly detection in federated learning via separable convolution and convergence acceleration

Abstract

The limitations of traditional centralized anomaly detection have led to the rise of distributed detection technology, which has become a hot topic in recent years. However, many key issues in distributed computing have attracted more attentions, such as large-scale data collection without exposing sensitive information and shared model exchange problems. Existing models have a large structure which have high requirements for edge device configuration. In this article, we propose a lightweight anomaly detection model based on federated learning, which ensures detection efficiency while protecting data privacy. In the local model, we incorporate depthwise separable convolution (DSC) into the Transformer to enhance both local and global information exchange within the window, and remove unnecessary modules to reduce the number of network structure parameters while ensuring accuracy. In the global model, we introduce a new parameter to the local client objective function to adjust the local training direction, which accelerates convergence speed and reduce client–server communication. Finally, we conducted comparative experiments on several public datasets, and the experimental results showed that our model ensures detection accuracy on par with the baseline while significantly reducing local optimization parameters. Meanwhile, after global optimization, the number of rounds required for model convergence is significantly reduced, reducing communication volume while ensuring detection accuracy. In general, the performance of our model was significantly improved compared to baseline.