![]() These three states are switched randomly in each batch. In order to alleviate the above shortcomings, Jia introduced a noise layer called Mini-Batch of Real and Simulated JPEG compression which had three states including simulated JPEG (similar to HiDDen ), real JPEG (to only train the decoder) and identity (no compression). This two-stage framework can bypass the non-differentiable process to train the network, but the disadvantage is that only the decoder can learn the features of compression. Thus, the decoder will learn to how to extract the watermark from a compressed image. In the second stage, the watermarked image will be compressed by real JPEG, and sent to the decoder. In the first stage, both encoder and decoder are trained without JPEG compression. Liu introduced a two-stage separable deep learning framework for image watermarking. In order to approximate the rounding calculation which is not differentiable during JPEG quantizing the DCT coefficients, StegaStamp used a piecewise function to replace the quantization steps around zero. For image compression, HiDDen used the discrete cosine transform (DCT) to build a differentiable compression layer as the approximation of compression algorithms. In this regard, existing algorithms propose some solutions. On the other hand, for end-to-end steganography, the way to resist compression is also a key issue. Jia also achieved an end-to-end video steganography with robustness to H.264 compression in which the encoder applied both motion synthesis and temporal correlation to maintain the visual quality. However, this method didn’t mention the robustness to video compression. The PyraGAN proposed by Chai introduces coding unit (CU) masks as an input of the encoder which facilitates the network to learn the texture complexity of the video content. The stego video generated by RIVAGAN has higher visual quality and higher a extraction rate, but it can only resist one type of video compression which is not the mainstream video compression format. RIVAGAN was another end-to-end video steganography model based on GAN, in which the attention module was added to help encoder and decoder extract features better. DVMark designed a video steganography network with a differentiable distortion layer to improve robustness, and the steganographic video (abbreviated as stego video) generated by DVMark can resist H.264 compression. Based on this, Wang developed a CNN to hide an inter-frame residual into a cover video. ![]() Weng found that hiding highly sparse data is significantly easier than hiding the original frames. However, the research of a video steganography algorithm based on deep learning is inchoate, and the number of relevant studies is very limited. In particular, videos have more redundancy compared with images which is beneficial for hiding more information. However, we know that the optional covers of steganography are much more than diverse ever before, including PDF files, pictures posted on social media applications, and videos played on websites. On the other hand, the first end-to-end image steganography model called HiDDeN has achieved very high payloads and good visual quality, and the experimental results showed that the end-to-end steganography model based on deep learning can achieve better performance compared with traditional steganography methods. On the one hand, compared with traditional methods, an end-to-end steganography model based on deep learning reduces the artificial feature extraction process, and allows neural networks to extract features and complete secret information embedding. This approach is often referred to as an end-to-end model. The decoder is responsible for extracting the secret information from the steganographic carrier. The encoder embeds the secret information into carrier and output the steganographic carrier. Steganography based on deep learning usually consists of two modules, an encoder and a decoder. However, with the rapid development of deep learning, more and more researchers attempt to apply this technology to steganography. ![]() ![]() Traditional steganography usually needs to apply relevant expertise to complete the embedding of secret information by modifying the carrier and redundant information in the coding process.
0 Comments
Leave a Reply. |