Detection of Virtual Private Network Traffic Using Machine Learning

The detection of unauthorized users can be problematic for techniques that are available at present if the nefarious actors are using identity hiding tools such as anonymising proxies or virtual private networks (VPNs). This work presents computational models to address the limitations currently experienced in detecting VPN traffic. A model to detect usage of VPNs was developed using a multi-layered perceptron neural network that was trained using flow statistics data found in the transmission control protocol (TCP) header of captured network packets. Validation testing showed that the presented models are capable of classifying network traffic in a binary manner as direct (originating directly from a user's own device) or indirect (makes use of identity and location hiding features of VPNs) with high degrees of accuracy. The experiments conducted to classify OpenVPN usage found that the neural network was able to correctly identify the VPN traffic with an overall accuracy of 93.71%. The further work done to classify Stunnel OpenVPN usage found that the Neural Network was able to correctly identify VPN traffic with an overall accuracy of 97.82% accuracy when using 10-fold cross validation. This final experiment also provided an observation of 3 different validation techniques and the different accuracy results obtained. These results demonstrate a significant advancement in the detection of unauthorised user access with evidence showing that there could be further advances for research in this field particularly in the application of business security where the detection of VPN usage is important to an organization.