Zero-Day Traffic Identification Using One-Dimension Convolutional Neural Networks And Auto Encoder Machine

Network traffic identification plays an important role in traffic engineering, anomaly detection and traffic billing. Recently, the machine learning and deep learning based algorithms have made a great success in identifying the known applications, where the training set and the test set are supposed to contain the same traffic classes. However, in a realistic scenario, the network traffic classifier may suffer from a low identification accuracy due to the substantial zero-day (unknown) traffic. The essential to solve this problem is to find the boundary between the known and zero-day traffic, which has not been well studied before. In this paper, based on the fact that for an Auto Encoder (AE) machine, the reconstruction error of a zero-day class is generally larger than that of a known class, we propose a zero-day traffic identification method using one-Dimension Convolutional Neural Networks (1D-CNN) and AE machine. In order to further improve the identification accuracy, we propose an algorithm to estimate the confidence possibility of the identification results based on the Extreme Value Theory. The experiments conducted on realistic traffic datasets demonstrate that our method has a great improvement in identification accuracy than the benchmarks.