Dual branch parallel steganographic framework based on multi-scale distillation in framelet domain

Abstract

The intuition behind image steganography is to hide secret image into cover image, thereby analysis and retrieval are avoided. Image steganography methods based on deep learning have recently gained popularity in comparison to traditional methods. However, most steganography methods have poor perceptibility and restoration robustness, resulting in weak quality of stego images and revealed images. To address this gap, we propose DBPSNet, an original dual branch parallel steganographic network based on knowledge distillation in framelet domain. Wavelet frame transform performed on cover image is used to migrate the steganographic framework in spatial domain to framelet domain. Afterwards, we establish the steganographic framework based on knowledge distillation, where the teacher branch produces feature maps layer-by-layer to guide student branch in learning how to generate reliable and realistic stego image. In addition, the random noise layer is used in reconstruction network, which helps improve the robustness of model by attempting to reveal image from the noisy stego image. Extensive controlled experiments on BOSSBase show that DBPSNet is ahead of some mainstream image steganography methods in terms of steganography and reconstruction effects, algorithm security and model robustness.