A Steganographic Scheme based on Perceptual Loss and Feature Fusion

The purpose of steganography is to convey information in public that no one else pays attention to. With the development of communication technique, the communication in the public channels also causes security problems.This is why the study of steganography has been concerned. Traditional steganographic algorithms usually extract the embedded information from the image itself. In this paper, a steganographic algorithm is proposed, which uses the fusion of feature maps in the higher layers of convolutional neural network (CNN) to embed information and form the corresponding stego image through network back propagation. The stego image is visually closer to the cover image through the designed mixed loss function. Compared with the pixel-wise loss function one, this algorithm is not easy to detect visually, and can resist certain disturbances in information reconstruction, which means it has good robustness.