General Network Traffic Anomaly Detection Method based on Large Language Models

Deep learning methods, while demonstrating strong performance in network traffic anomaly detection, often require large-scale training datasets and lack generalization ability. To address these issues, utilizing the powerful information extraction and reasoning capabilities of large language models, we propose a network traffic time series anomaly detection model with high generalizability. We design a text prototype construction method based on data reconstruction, to obtain a text prototype more suitable for the time series field; and we design a set of text templates, which provides effective information such as global data information, expert knowledge for anomaly detection tasks in large language models. Experiments show that our model achieves the best performance in anomaly detection and generalization compared to other methods.