Lightweight Deep Learning-Based Network Traffic Classification Using Explainable Artificial Intelligence

To address the growing demand for internet usage and connected devices, it is essential to implement robust Network Traffic Classification (NTC) methods that guarantee network security and optimised resource utilisation. Deep learning approaches are continuously gaining traction due to their capabilities in learning complex network traffic patterns and effective classification. However, the performance of deep learning (DL) approaches comes with the cost of huge computational resources that struggle to perform under a resource-constrained environment. Therefore, this study proposes enhancing DL-based NTC by integrating an Explainable Artificial Intelligence (XAI) technique, the Shapley Additive exPlanations (SHAP), to identify and select the most important features. This study leverages the explainability of XAI to enhance model performance and optimise resource utilisation through feature selection. To evaluate the selected features, this paper compares three algorithms, which are Convolutional Neural Networks (CNN), Multi-Layer Perceptron (MLP), and Gated Recurrent Units (GRU), using the ISCXVPN2016 dataset on an NVIDIA Tesla T4. The result shows a $\mathbf{1 2 . 4 2 \%}, \mathbf{1 0 . 7 2 \%}$, and $\mathbf{1 0 . 2 3 \%}$ decrease in RAM usage for MLP, CNN, and GRU, respectively, while reducing the GPU memory usage by $6 \%$ and an improvement up to $4.99 \%$ F1-score compared to the state-of-the-art.