The Research and Analysis of Robust Image Steganography for Arbitrary Style Transfer
Boning Zhang, Dongjun Tan, Wujian Ye, Xueke Zhu, Yijun Liu · 2024
The use of deep learning technology has notably improved the performance of image steganography. However, current methods are lacking in stylized environments and require retraining of the steganography model for each specific style, resulting in high training costs, limited applications, and low effectiveness in anti-steganalysis. To address these issues, a robust method for image steganography in arbitrary style transfer is proposed. This method introduces an arbitrary stylization model and a residual module with enhanced structural loss. Initially, an image preprocessing module aligns sizes and fuses features of the input cover image and secret image. Subsequently, the feature extraction ability of the encoder-decoder-based steganography model is further enhanced by the introduction of two-layer residual blocks. Finally, the attention-based arbitrary stylization network and structural loss are utilized to optimize the network training of the steganography model, achieving effective embedding and decryption of the secret image in the context of arbitrary stylization. Experimental analysis demonstrates that the proposed method can achieve an SSIM of 0.91, a PSNR of 28.11, and a BER of 0.07 between the input secret image and the decrypted output. The stylized steganographic (stego) images exhibit improved visual quality and are visually more difficult to detect for embedded traces compared to the original stego image. The use of the YeNet model for steganalysis on stylized stego images yields a detection accuracy as low as 42.33%, about 26% lower than before stylization, significantly reducing the risk of discovering steganographic information. Therefore, the proposed method can achieve more effective steganography for arbitrary stylization, with enhanced scalability, imperceptibility, and anti-steganalysis capabilities.