YouTube Video Identification in Encrypted Network Traffic —A Case of Attacking Assumptions

Identifying the video in the network traffic is in fact a tough challenge for researchers. Recent studies on video identification in the network traffic use many different assumptions that increase the difficulty of practical deployment of the techniques. This study proposes an attack using a Convolutional Neural Network for video identification in the network traffic and also targets assumptions to understand their effects on the accuracy of the attack. Two basic assumptions are targeted in this work: (1) Time of starting the videos at the client and previous samples in the dataset used for training of models should be same and (2) all normalization techniques for normalizing features before feeding them to the classifier will work equally. The results show that even if the video starts playing from any random position, the convolutional neural network can identify the video with high accuracy of 95%. There is no restriction of the exact match between the start time of the video in the client and in the samples used for training. Moreover, max normalization technique provides better accuracy with the model than sigmoid function.