SSCMT-ETC: A Semi-Supervised Contrastive Mean Teacher Model for Encrypted Traffic Classification

The openness and diversity of cyberspace have led to the widespread adoption of traffic encryption technology that can protect privacy and security. There is a urgency to efficiently classify encrypted traffic to identify and deal with potential network security threats. Although deep learning (DL) -based solutions have emerged and proven effective for this task. Due to the dynamism and mobility of networks, their reliance on labeled data renders them unsuitable for the situation where new encryption services are constantly emerging. Therefore, finding ways to make use of abundant unlabeled data for enhancing classification performance in the case of scarce labeled data is a priority. Benefiting from the advantages of both semi-supervised learning (SSL) and contrastive learning (CL) in leveraging unlabeled data, we put forward a classification approach that combines the two and is suitable for limited labeled samples. Specifically, The SSL framework based on mean teacher (MT) exploit both labeled and unlabeled data through the introduction of consistency loss. Contrastive loss, serving as an additional unsupervised constraint apart from consistency loss, assists the encoder in discovering more varied representations. Experiments conducted on the public dataset ISCXVPN-NonVPN2016 validate the competence of our approach.