Video steganography network based on 3DCNN

In recent years, the steganography scheme based on neural network has made many significant progress on images, but it is still in the exploratory stage in the field of video steganography. By using long skip connections to extract the spatio-temporal information in the video, this paper proposes a 3DCNN full-video steganography network. The network takes a pair of cover and secret video sequences as input, and uses a stego network to output a spatio-temporal residual sequence, which is added to the cover video as a small disturbance. A video classification network is proposed, which can be used to identify the cover video frame and the stego video frame to assist the message receiver to extract the secret message correctly. We chose UCF101 video data set as the training and testing set of the network model. We used various video quality evaluation indicators (PSNR, SSIM, Pixel distribution) to measure the performance evaluation of the stego video network, and proved the anti-detection of the stego video by using some stego detection algorithms. Under the training and testing of the data set of stego videos generated by the stego network, the classification accuracy of the proposed video classification network reaches about 93%.