Deep Photographic Steganography for Display-to-Camera Communication

Intricate and diverse image distortions exist in display-to-camera communication, a technology that displays steganographic images on a screen, then decodes convert messages by a camera. Modeling these distortions has always been a difficult problem. Most of the previous work is to use prior knowledge to artificially design the noise layer to approximate the real situation, however, it is of less authenticity. In this paper a more impeccable channel distortion model based on generative adversarial network, named D2C-Distortion is established to simulate the display-to-camera transmission channel caused by complex image distortions, based on which, a complete set of end-to-end photographic steganography scheme consisting of encoding, channel distortion and decoding networks are proposed. Experimental results show that the scheme with D2C-Distortion process is more robust against display-to-camera distortions compared with the one without it, demonstrating the effectiveness of the proposed distortion model.