A Cost-Sensitive Deep Learning-Based Approach for Network Traffic Classification

Network traffic classification (NTC) plays an important role in cyber security and network performance, for example intrusion detection facilitating a higher quality of service. However, due to the unbalanced nature datasets, NTC can be extremely challenging poor management degrade performance. While existing methods seek re-balance data distribution through resampling strategies, such approaches are known suffer from information loss, overfitting, increased model complexity. To address these challenges, we propose new cost-sensitive deep learning approach increase robustness classifiers against imbalanced class problem NTC. First, dataset is divided into different partitions, cost matrix created each partition by considering distribution. Then, costs applied function layer penalize errors. In our approach, diverse type misclassification because specifically generated partition. determine its utility, implement proposed method two classifiers, namely: stacked autoencoder convolution neural networks. Our experiments on ISCX VPN-nonVPN show that obtain performance low-frequency classes, comparison three other methods.