<![CDATA[Deep neural networks are highly sensitive to subtle pixel-level variations, a property that is commonly discussed in the context of adversarial learning. Rather than treating this characteristic solely as a vulnerability, this study explores its potential for image data hiding. This paper proposes StegaFlow, an end-to-end deep learning-based framework for image steganography and blind watermarking, in which encoder and decoder networks are jointly optimized to embed binary messages into digital images while preserving visual quality and enabling reliable message recovery. To improve robustness against real-world image degradations, the training process incorporates multiple distortion models, including Gaussian blur, pixel-wise dropout, image cropping, and JPEG compression. Because JPEG compression is inherently non-differentiable, differentiable approximations are introduced during training to enable effective gradient-based optimization. Experimental results evaluate the proposed framework in terms of capacity, secrecy, and robustness, and show that it achieves strong performance under common image distortions. In particular, the framework demonstrates improved robustness to several spatial-domain degradations, while adversarial training contributes to better perceptual quality by reducing visible embedding artifacts. These findings indicate that end-to-end neural optimization provides a flexible and effective approach for robust image data hiding]]>
Copyrights © 2026
