Optimized DenseNet Architecture for Efficient Classification of Encrypted Internet Traffic

Abstract

The increasing reliance on Internet-based services has rendered secure and efficient network traffic classification critical. Conventional methods for categorising traffic, such as port and payload methods, often struggle with the challenges posed by encrypted traffic. Deep learning techniques have emerged as a predominant method for traffic classification given their success in domains such as image recognition, document analysis, and genomics. This research proposes an enhanced DenseNet architecture that leverages deep learning to accurately classify encrypted Internet traffic categories. This approach introduces a compression layer into the DenseNet architecture to address the co-adaptation problem as a result of the information flow and optimise the accuracy of the CNN. An Intrusion detection dataset from the Canadian Institute of Cybersecurity was used to evaluate the architecture. The optimised DenseNet architecture was evaluated using metrics such as precision, recall, accuracy, F1-Score, False Positive Rate and Area under the ROC Curve. Experimental results show that the approach can distinguish various encrypted Internet traffic categories.