TrafficMAE: A network traffic classification model based on masked autoencoder

Network traffic classification is critical for security and performance. Encrypted protocols challenge traditional methods, and supervised approaches suffer from limited labeled data. We propose TrafficMAE, a masked autoencoder using a Session Video representation to encode temporal and hierarchical traffic information. Our self-supervised pre-training combines masked reconstruction with bidirectional packet direction prediction to learn from unlabeled data. Fine-tuning fuses statistical and learned features. Evaluations on five tasks show TrafficMAE outperforms existing methods with superior F1-scores, strong few-shot performance, and better efficiency compared to transformer-based baselines.