Self-Attention based Traffic Anomaly Detection for Cloud Servers using Line Chart Patch Filling

With the rapid expansion of cloud servers in applications, detecting anomalies in network traffic accessing cloud servers has become a core concern of cloud server providers. Traditional anomaly detection methods suffer from low accuracy and inefficiency, thus failing to meet the demands. Meanwhile, deep learning techniques struggle with the issues of data imbalance. To address these challenges, we propose Line Chart Patch Filling assisted Self Attention (LCPF-SA) mode. LCPF-SA first transforms raw time series data into images by producing line chart, filling parts of the line chart with selected patterns. The images are then classified by using a multi-layer model that integrates Convolutional Neural Network (CNN) and Self-Attention (SA) mechanism. To facilitate the training and evaluation of network traffic anomaly detection algorithms, we construct a dedicated dataset and named HQUNetTraffic dataset. The performance of LCPF-SA is comparatively evaluated on our dataset and a public ElectricDevices dataset from UCR. Experiment results show that LCPF-SA outperforms existing anomaly detection techniques and exhibits greater robustness to data imbalance.