Universal video steganalysis in HEVC intra frame based on the convolutional residual network

High Efficiency Video Coding (HEVC) offers a variety of suitable carriers for video steganography, contributing to the rapid advancement of the field. However, most of the existing steganalysis algorithms can only detect a single type of video steganography and are unable to detect other types of video steganography. To detect different video steganography methods in multiple domains of intra frame, a universal video steganalysis algorithm based on convolutional residual networks is proposed in this paper. Firstly, we introduce and analyze three types of video steganography in intra frame: video steganography based on intra partition structures, transform unit (TU) partition modes and intra-prediction modes (IPM). It reveals that these types of algorithms have minimal impact on the spatial domain of the video. However, they may disrupting the optimality of the local structure or the intra-prediction modes in specific regions of the video. Building on this observation, we introduce the concept of the IPM map, which not only effectively captures the IPM information of the video frame but also partially reflects the partition structure information of the video. Then, the IPM maps from both the original and recompressed video frame can be extracted. The difference between these two maps serves as the input to the convolutional residual network. Finally, a convolutional residual network, called IFSN (Intra Frame Steganalysis Network), comprising preprocessing, feature extraction, feature representation and binary classifier, is used for training and detecting the video frames. Experimental results demonstrate that the proposed convolutional residual network effectively detects different video steganography algorithms in multiple domains of intra frame.