Reduce memory consumption for internet traffic classification

Abstract

Application level traffic classification is an essential requirement for stable network operation and resource management. However, the classification's processing tends to face low resources when high volumes of traffic are being classified in high-speed networks in real time. Memory consumption considered to be a serious issue during classification processing time. In this paper, a data reduction method is proposed to decrease redundant data entry during the preprocessing phase with regard to accuracy classification. The proposed active build-model random forest (ABRF) eliminates redundant data-entry by utilising feature selection algorithm during the preprocessing phase. The proposed system successfully reduces the memory space of the entire classification process. The system is evaluated by comparing the proposed system against four classifiers (RF, NB, SVM and C5.0) and four features selection techniques (FCBF, SFE, Chi2 and GR). DR reported excellent results amongst the NB, C5.0 and RF. The results were optimised due to the data excluding 314,216 out of 774,013. Moreover, C5.0 consumed less memory space due to the decreased depth of C5.0 tree model. In conclusion, the DR was most effective on the RF model due to the nature of the ensemble classifier.