Reducing Blocking Artifacts in CNN-Based Image Steganography by Additional Loss Functions

• Published in: 2020 12th International Conference on Knowledge and Systems Engineering (KSE) pp. 61 - 66
• Main Authors: Dung Pham, Tuan, Cuong Ta, Viet, Thanh Thuy Pham, Thi, Ha Le, Thanh
• Format: Conference Proceeding
• Language: English
• Published: IEEE 12.11.2020
• Subjects: blocking artifact; convolution network; image steganography
• DOI: 10.1109/KSE50997.2020.9287408

Our work improves the encoded image quality from HiDDeN framework, an end-to-end image steganography based on deep convolution neural network. In the encoding phase of HiDDeN framework, to embed a message in a cover image, it is required to split the cover image into smaller image blocks and embed the message bits in each block in parallel. These embedded blocks are then combined to form an encoded image that has the same size as the cover image. This image reconstruction process causes artifacts that appear on the boundaries of the blocks. This can be explained by the fact that when message bits are embedded in the image blocks, the pixel-level information of each image block is unequally alternated. In order to reduce block artifacts, in this work we propose a blocking loss as an additional objective function in HiDDeN framework. This loss measures the difference between encoded images and modified versions of the cover images. The proposed method is evaluated on COCO 2014 and BOSS datasets and the experimental results show the effectiveness in reducing the block artifacts that appeared in the encoded images of HiDDeN framework. This has an important impact on increasing the invisibility or transparency of the steganography system. In addition, the experimental result on secrecy of the proposed method also indicates the same performance as the HiDDeN pipeline.