| Title |
Enhanced BERT for Encrypted Traffic Classification: Lightweight and Robust Approach |
| Authors |
조민호(Minho Cho) ; 권용석(Yongseok Kwon) ; 안세영(Seyoung Ahn) ; 조성현(Sunghyun Cho) |
| DOI |
https://doi.org/10.5573/ieie.2024.61.11.49 |
| Keywords |
BERT; Deep learning; Encrypted traffic classification; Lightweight; Security |
| Abstract |
This paper proposes SRB-ET (Slim and Robust BERT for Encrypted Traffic), a lightweight and robust BERT model designed for encrypted traffic classification in resource-constrained network systems. SRB-ET employs neighboring weight averaging and half-probability label prediction techniques to reduce model size while maintaining high classification performance. The neighboring weight averaging technique minimizes loss and enhances performance by averaging the weights of pruned layers with adjacent layers. The half-probability label prediction technique efficiently learns category features by predicting labels with a 50% probability during pre-training, enabling faster convergence. Experimental results using the ISCX VPN-nonVPN dataset demonstrate that SRB-ET maintains traffic classification accuracy while reducing the number of parameters by 15.12%, decreasing inference time by 45.1%, and improving training speed by 54.9% compared to existing methods. |