Securing Data Via Vision Transformer with Residual Learning in Image Steganography

Image steganography, a subfield of data hiding, involves embedding secret information into digital media such as images. Despite extensive research, challenges persist in balancing visual imperceptibility, payload capacity, and robustness. Existing methods, including traditional algorithms and convolutional neural networks (CNNs), often suffer from limited global context awareness, suboptimal feature extraction, and noticeable image distortion when embedding high payloads. To address these gaps, this study proposes a Vision Transformer-based steganography framework that leverages enhanced residual learning for high-fidelity message embedding. Specifically, dual-stream architecture incorporates self-attention mechanisms to capture global dependencies and refined residual pathways to preserve local texture, thereby enabling the model to embed larger payloads while maintaining high visual quality. Experimental results demonstrate the effectiveness of the method, achieving a peak signal-to-noise ratio (PSNR) of 46.48 dB and a structural similarity index measure (SSIM) of 0.9982, outperforming state-of-the-art techniques. Furthermore, the proposed framework shows strong robustness, with a minimal entropy deviation of 0.0002.