Work

Timely Classification of Encrypted or ProtocolObfuscated Internet Traffic Using Statistical Methods

Year: 2023
Type: dissertation
Abstract: Internet traffic classification aims to identify the type of application or protocol that gene­rated a particular packet or stream of packets on the network. Through traffic classifi­cation, Internet Service Providers (ISPs), governments, and network administrators can access basic functions and several solutions, including network management, advanced network monitoring, network auditing, and anomaly detection. Traffic classification is essential as it ensures the Quality of Service (QoS) of the network, as well as allowing efficient resource planning. With the increase of encrypted or obfuscated protocol traffic on the Internet and multi­layer data encapsulation, some classical classification methods have lost interest from the scientific community. The limitations of traditional classification methods based on port numbers and payload inspection to classify encrypted or obfuscated Internet traffic have led to significant research efforts focused on Machine Learning (ML) based classification approaches using statistical features from the transport layer. In an attempt to increase classification performance, Machine Learning strategies have gained interest from the sci­entific community and have shown promise in the future of traffic classification, specially to recognize encrypted traffic. However, ML approach also has its own limitations, as some of these methods have a high computational resource consumption, which limits their application when classifying large traffic or real­time flows. Limitations of ML application have led to the investigation of alternative approaches, including feature­based procedures and statistical methods. In this sense, statistical analysis methods, such as distances and divergences, have been used to classify traffic in large flows and in real­time. The main objective of statistical distance is to differentiate flows and find a pattern in traffic characteristics through statistical properties, which enable classification. Diver­gences are functional expressions often related to information theory, which measure the degree of discrepancy between any two distributions. This thesis focuses on proposing a new methodological approach to classify encrypted or obfuscated Internet traffic based on statistical methods that enable the evaluation of network traffic classification performance, including the use of computational resources in terms of CPU and memory. A set of traffic classifiers based on Kullback­Leibler and Jensen­Shannon divergences, and Euclidean, Hellinger, Bhattacharyya, and Wootters dis­tances were proposed. The following are the four main contributions to the advancement of scientific knowledge reported in this thesis. First, an extensive literature review on the classification of encrypted and obfuscated Internet traffic was conducted. The results suggest that port­-based and payload­-based methods are becoming obsolete due to the increasing use of traffic encryption and multi­layer data encapsulation. ML­based methods are also becoming limited due to their com­putational complexity. As an alternative, Support Vector Machine (SVM), which is also an ML method, and the Kolmogorov­Smirnov and Chi­squared tests can be used as refe­rence for statistical classification. In parallel, the possibility of using statistical methods for Internet traffic classification has emerged in the literature, with the potential of good results in classification without the need of large computational resources. The potential statistical methods are Euclidean Distance, Hellinger Distance, Bhattacharyya Distance, Wootters Distance, as well as Kullback­Leibler (KL) and Jensen­Shannon divergences. Second, we present a proposal and implementation of a classifier based on SVM for P2P multimedia traffic, comparing the results with Kolmogorov­Smirnov (KS) and Chi­squaretests. The results suggest that SVM classification with Linear kernel leads to a better clas­sification performance than KS and Chi­square tests, depending on the value assigned to the Self C parameter. The SVM method with Linear kernel and suitable values for the Self C parameter may be a good choice to identify encrypted P2P multimedia traffic on the Internet. Third, we present a proposal and implementation of two classifiers based on KL Diver­gence and Euclidean Distance, which are compared to SVM with Linear kernel, configured with the standard Self C parameter, showing a reduced ability to classify flows based solely on packet sizes compared to KL and Euclidean Distance methods. KL and Eucli­dean methods were able to classify all tested applications, particularly streaming and P2P, where for almost all cases they efficiently identified them with high accuracy, with re­duced consumption of computational resources. Based on the obtained results, it can be concluded that KL and Euclidean Distance methods are an alternative to SVM, as thesestatistical approaches can operate in real­time and do not require retraining every time a new type of traffic emerges. Fourth, we present a proposal and implementation of a set of classifiers for encrypted Internet traffic, based on Jensen­Shannon Divergence and Hellinger, Bhattacharyya, and Wootters Distances, with their respective results compared to those obtained with metho­ds based on Euclidean Distance, KL, KS, and Chi­Square. Additionally, we present a com­parative qualitative analysis of the tested methods based on Kappa values and Receiver Operating Characteristic (ROC) curves. The results suggest average accuracy values above 90% for all statistical methods, classified as ”almost perfect reliability” in terms of Kappa values, with the exception of KS. This result indicates that these methods are viable op­tions to classify encrypted Internet traffic, especially Hellinger Distance, which showed the best Kappa values compared to other classifiers. We conclude that the considered statistical methods can be accurate and cost­effective in terms of computational resource consumption to classify network traffic. Our approach was based on the classification of Internet network traffic, focusing on sta­tistical distances and divergences. We have shown that it is possible to classify and ob­tain good results with statistical methods, balancing classification performance and the use of computational resources in terms of CPU and memory. The validation of the pro­posal supports the argument of this thesis, which proposes the implementation of statis­tical methods as a viable alternative to Internet traffic classification compared to methods based on port numbers, payload inspection, and ML.
Source: HAL
Language: English
Open Access status: green
OpenAlex ID: W4392276943
Cites: 209
Cited by:
Related to: 10