Deep Learning-based Quantitative Steganalysis to Detect Motion Vector Embedding of HEVC Videos

摘要

Generally, the purpose of a steganalysis algorithm is to establish the presence of secret messages in the stego data. However, quantitative steganalyzers can reveal more information about the secret communication by estimating the exact volume of embedded messages. Quantitative steganalysis is a crucial step for breaking secret codes in many practical scenarios. This work concerns about the quantitative steganalysis of videos. Most video steganographical algorithms embed secret messages by modifying the values of motion vector in the compressed domain. We propose a general framework for constructing video quantitative steganalyzers that are able to detect the embedding of motion vectors based on features learned by deep convolutional neural networks. Considering that video structure is quite different from that of image, we focus on the construction of input data matrix for deep convolutional neural network and the robustness of the detection network against different bitrates. Because videos at different embedding rates have different steganalysis features, we use multiple models to extract features for the construction of feature vector. Experimental results have validated the proposed method. Our deep learning-based steganalyzer obtained satisfactory estimation accuracy on testing HEVC videos at multiple embedding rates under different video bitrates.