A Biologically-Inspired Approach to Network Traffic Classification for Resource-Constrained Systems

Abstract

Internet traffic classification has been studied widely in recent years, and many machine learning approaches have been applied to it. Internet traffic classification has increased in relevance in recent years because of its potential applications in the business world. Information about network traffic has many benefits in network design, security, management and accounting. Internet traffic classification is especially important to the adaptive networks often used in cloud computing, which must use data gleaned from the network to adjust to network conditions on-the-fly. This information is most easily collated from the huge amount of information going through a modern network with machine learning algorithms, which adjust themselves to the conditions of the network. In previous research, Artificial Immune System (AIS) algorithms have been used to classify malicious and benign network traffic in support of Intrusion Detection Systems [1]. Because of their versatility and their low sensitivity to the values of the input parameters, we are motivated to explore the value of using AIS inspired algorithms in support of flow-based traffic classification. In this paper, we propose an AIS inspired algorithm for flow-based traffic classification, and evaluate its performance with and without the use of kernel functions. We utilize a publicly-available dataset to compare our results with other approaches that have been proposed in the recent literature. We provide several measures of the classification performance of the algorithm, as well as share our experience on the best features of the algorithm for this particular application. We also evaluate the proposed algorithm, comparing it with two other classification algorithms, and draw conclusions based on our findings. The algorithm generalizes well and gives high accuracy even with a small training set when compared to other algorithms, although the training and classification times were higher. The algorithm is also insensitive to the values of the input parameters, which makes it attractive for embedded and Internet of Things applications. The research presented here is a longer exposition of the work in [2].