FSAM Framework for Online CDN-Based Website Classification

With the development of the Internet, CDN has become an important infrastructure of the Internet by hosting websites or certain components. To optimize the QoS of CDN and user experience, it is necessary to classify the website or service traffic on CDN to set different transmission priorities for different categories of websites. However, few papers or datasets are involved in this direction. Therefore, we classify website categories and CDN brands based on encrypted traffic using deep learning-based methods, which are the mainstream methods in current research. In this paper, we first collect traffic from various categories of websites (including self-built ones) hosted on several CDN brands on different devices. The high-quality, strongly related, interference-free, and representative dataset WebT2023 is released to the public via Google Drive. On this basis, a new end-to-end framework Flow Sampling Attention Model(FSAM) for CDN website classification is proposed. Several comparative experiments under different models are completed, in which the selection of hyperparameters is analyzed. The experimental results show that FSAM outperforms the state-of-the-art model in both performance and efficiency, improving the accuracy by 5.2% in website category classification, 9.2% in CDN brand classification, and efficiency over 10 times by flow sampling strategy.