METC-MVAE: Mobile Encrypted Traffic Classification With Masked Variational Autoencoders
Wei Cai, Zhen Li, Peipei Fu, Chengshang Hou, Gang Xiong, Gaopeng Gou · 2022
User privacy and information security have received significant attention in recent years, and the encryption rate for mobile traffic has risen, posing significant difficulties to traditional traffic classification methods. Machine learning based methods have become popular for addressing this issue. However, it relies on manual features a lot to extract hidden pattern. Deep learning based approaches can learn features from raw traffic automatically, but they need labeled data, and the size of the dataset has a direct effect on the modeling effect. This paper proposes a self-supervised mobile encrypted traffic classification method to address these issues. called METC-MVAE. It learns a relational representation of traffic from large-scale unlabeled data, and can fast learn the downstream task with a small dataset by finetuning. We test our method on both private and public datasets. We collected more than 1,160,000 traffic flows from 300 of the most popular mobile encryption Apps as the private dataset. On three downstream tasks, our method is better than state-of-the-art approaches by 6.7%, 2.6%, and 6.9%.