Adversarial Training for Improved VPN Traffic Classification Using EfficientNet-B0 and Projected Gradient Descent

This paper addresses the challenges of VPN/Non-VPN classification in increasingly complex network traffic by proposing an adversarial training-based classification method.The model integrates the EfficientNet-B0 architecture with a biLSTM structure to handle variable-length packet sequences, transforming them into fixed-size feature vectors for further extraction.Two fully connected layers increase the feature dimensions, preparing the data for convolutional processing.By incorporating Projected Gradient Descent adversarial training, the model is finetuned for robust classification against adversarial attacks, such as packet delays and congestion, while maintaining strong performance in normal traffic.Experimental results show high classification accuracy, achieving 99.81% on normal traffic and 99.35% on adversarial traffic using the ISCX 2016 dataset, with minimal trade-offs.This approach offers a scalable and resilient solution to VPN/Non-VPN classification in hostile network environments.