Supervised Anomaly Detection in Computer Network Traffic

Two of the most challenging tasks in computer network are traffic classification and outlier detection. Due to increasing network diversity and growth of network applications some of the most widely-used approaches such as port-based classification are not sufficiently accurate anymore. In the past years new proposed machine learning methods achieved good results for these problems. This work addresses the problems of network traffic classification and outlier detection. Our proposed approach is in the direction of Active Learning. A combination of Convolutional Neural Network, Autoencoder and Principal Component Analysis intends to decide whether packets are from already known and labeled application or from an unknown. The models can be then retrained with a larger dataset. The system is evaluated with our own dataset and its results show promising performance.