Structured 2D Data Transformations for Resource-Efficient and Explainable CNN-Based Encrypted Video Traffic Classification

The widespread adoption of modern cryptographic protocols has rendered traditional packet payload inspection obsolete, pushing network management toward deep learning-based behavioral profiling. While Convolutional Neural Networks (CNNs) have emerged as a powerful tool for classifying encrypted video traffic, a systematic understanding of how different 1D-to-2D data transformation schemes affect classification accuracy, computational complexity, and model interpretability is still lacking. In this study, the author presents a comparison of nine data transformation schemes divided into mathematical encodings and statistical feature–time representations, evaluated under a standardized CNN pipeline and benchmarked against Large Vision Models, including ResNet-50 and Vision Transformers (ViTs). Using real encrypted video traces, the results show that structured representations, such as Feature-Grouped and Multi-Scale matrices, achieve high fine-grained classification performance, outperforming mathematical encodings such as Gramian Angular Fields while reducing training time by 21.4×–23.9×. Explainability analysis via Gradient-weighted Class Activation Mapping (Grad-CAM) and Attention-to-Burst ratios indicates that CNNs exploit both active download bursts and provider-specific silence intervals as discriminative cues in Dynamic Adaptive Streaming over HTTP (DASH) streaming, while ViTs suffer from a lack of spatial inductive bias, resulting in a performance degradation of up to 12 percentage points in Global F1-score and a 7.5× increase in training latency when trained from scratch. This paper concludes that structured feature–time representations provide the optimal trade-off for real-time, explainable edge network deployment.