어텐션을 활용한 BERT 기반 암호화 트래픽 분류 모델 분석 연구

Abstract

Despite the high performance in the encrypted traffic classification tasks, BERT-based models have the risk of error caused by the lack of explanation. In this paper, we implement a Bidirectional Encoder Representation from Transformer (BERT)-based traffic classification model and analyze the model by utilizing attention scores to determine which features influence the classification results. We show that the first 30 bytes of packets are dominant features for classifying traffic into their respective application.