Automated Encrypted Proxy Traffic Classification via Optimized 1D-CNN and Feature Attention

This study addresses the critical challenge of identifying encrypted proxy traffic, which poses significant risks to network security due to its strong anonymity and evolving protocols. Existing methods primarily rely on manual feature engineering and struggle with generalization across new proxy techniques, while the lack of comprehensive datasets further limits detection accuracy. To overcome these limitations, our study proposes an enhanced 1D Convolutional Neural Network (1D-CNN) model integrated with Particle Swarm Optimization (PSO) for hyperparameter tuning and a self-attention mechanism for feature refinement. The experiments first evaluate baseline 1D-CNN performance, followed by ablation studies to validate the contributions of PSO optimization, self-attention, and an improved loss function. Finally, classification tests are conducted on three datasets (binary, 3-class, and 5-class tasks) to assess generalization. The results demonstrate that our hybrid model achieves 98.4% accuracy, outperforming conventional methods by 12.7%, while reducing computational redundancy by 38%. This research establishes a robust framework for encrypted proxy traffic identification, offering significant improvements in detection precision and adaptability for real-world cybersecurity applications.