SmoteAdaNL: a learning method for network traffic classification

Abstract

Machine learning based network traffic classification is a critical technique for network management, and has attracted much attention. Recently, most of the researchers focus on achieving high flow classification accuracy (FCA). However the amount of ''mice'' flows is more than that of ''elephant'' flows in the Internet, these classifiers hence are more suitable for ''mice'' flows, but have low byte classification accuracy (BCA). To address this issue, the notion of byte misclassification is firstly explored. According to the exploration that most misclassified bytes belong to the minority class, a novel method of network traffic classification is proposed by combining the data re-sampling and ensemble learning algorithms. To enhance the classification accuracy of the minority class, the data re-sampling algorithm is employed to increase the number of minority class flows. The data re-sampling however will change the data distribution and degrade the generalization of a classifier. A boosting-style ensemble learning algorithm with the consideration of ensemble diversity hence is employed to improve the generalization. The experiments conducted on the real-world traffic datasets show that the proposed method achieves over 90 % BCA and 96 % FCA on average, and improves about 7.15 % BCA by comparing with the existing methods.

Providing secure and reliable communication for next generation networks in smart cities

“…Their work proved the effectiveness of heterogeneous models in comparison to homogeneous solutions. In [51], the authors presented a hybrid approach that uses resampling to maximize the flow in a minority class. It also applies the ensemble method to improve the classifier generalization.…”

Online Data Center Traffic Classification Based on Inter-Flow Correlations

“…• Algorithmic solutions [7], [9] are more complex as they require a complete understanding of the classifier's reasoning in order to be able to consider the possible options of algorithm enhancement [15]. Recently, ensemble algorithms have become one of the most deployed algorithm level strategies for data imbalance problem.…”

“…At the algorithm level, [9] used SMOTE to reduce the imbalances between minority and majority classes, combined with a boosting-like strategy with the aim of enhancing Byte Classification Accuracy (BCA). Along with the misclassification rate, this work computes for each sample a penalty term expressing ensemble diversity (low disagreement degree between classifiers).…”

“…Indeed, most of the prior work on mice and elephant flow identification [3], [8] either assume the uniformity of traffic classes and thus do not consider the data imbalance problem but rather focus on enhancing traffic classification performances in ideal scenarios; or they classify only majority classes, because there approaches lack of rebalancing strategies and consequently fail to classify minority classes. Other algorithms [9], can not provide early classification and instead build offline methods since their algorithms do not guarantee efficient classification with only a few samples. Therefore, the most critical problem in DC traffic classification is to perform online flow classification when traffic is highly imbalanced because packet loss or disorder resulting from using few packets in the online classification process (also called packet reduction) can affect the performances of the classifiers in real-time networks [8] and imbalanced datasets may result in inefficient resource provisioning (congestion or over provisioning).…”

Security Enrichment in Intrusion Detection System Using Classifier Ensemble

“…Experimentation carried out on various data sets of intrusion detection proves effectiveness of heterogeneous models compared with homogeneous models. Liu et al [9] presented a hybrid approach SmoteAdaNL that applies resampling in order to increase number of flows in minority class and then diversified ensemble technique to improve the generalization of classifier. Weight assignment to the misclassified flows helps to improve the classification performance.…”