On the impact of packet sampling on Skype traffic classification

Nowadays, traffic classification technology addresses the exciting challenge of dealing with ever-increasing network speeds, which implies more computational load especially when on-line classification is required, but avoiding to reduce classification accuracy. However, while the research community has proposed mechanisms to reduce load, such as packet sampling, the impact of these mechanisms on traffic classification has been only marginally studied. This paper addresses such study focusing on Skype application given its tremendous popularity and continuous expansion. Skype, unfortunately, is based on a proprietary design, and typically uses encryption mechanisms, making the study of statistical traffic characteristics and the use of Machine Learning techniques the only possible solution. Consequently, we have studied Skypeness, an open-source system that allows detecting Skype at multi-10Gb/s rates applying such statistical principles. We have assessed its performance applying different packet sampling rates and policies concluding that classification accuracy is significantly degraded when packet sampling is applied. Nevertheless, we propose a simple modification in Skypeness that lessens such degradation. This consists in scaling the measured packet interarrivals used to classify according to the sampling rate, which has resulted in a significant gain.