A Novel Anti-rounding Image Steganography Method for Improved UNet++

A Novel Anti-rounding Image Steganography Method for Improved UNet++

연구 분야: Strategies

학회: Chinese Conference on Pattern Recognition and Computer Vision (PRCV)

Deep hiding involves embedding a secret image within a cover image so that the secret becomes imperceptible to ensure the security of steganography. However, in practical applications, accuracy loss is due to data conversion during encoding and decoding. The encoder generates the tensor matrix during training, while the practical application corresponds to the image preservation extraction process; the error caused by the rounding operation in this process will affect the accuracy of the secret information extraction during decoding. Aiming at the currently existing methods ignoring the rounding error problem, this paper proposes a novel anti-rounding image steganography method for improved UNet++. We design an anti-rounding function (ARF) to simulate the real steganographic environment, training the model to overcome the non-differentiable rounding distortion problem. When rounding error is considered, the accuracy of the extracted image decreases. To ensure the image quality in real application scenarios, in this paper, we use the improved UNet++ to ensure the quality of steganographic images by cascading convolution and ECA attention mechanisms and propose the Multiscale Mixing Module (MSMM) for feature fusion. In the extraction stage, the corresponding cover image is selected through a shared cover image library, followed by processing the stego image to obtain a different image before feeding it to the extraction network to guarantee the steganography process’s safety. The experiment results indicate the method’s capability to enhance the imperceptibility of steganography significantly, and the steganography process has better security.