RobustStegFormer: A Cross-Attention Guided and Noise-Aware Transformer for Image Steganography

Image steganography embeds secret information into digital images while preserving visual imperceptibility and data security. Although deep learning-based methods have improved embedding capacity and reconstruction quality, their robustness against real-world distortions such as noise and compression remains limited. We propose RobustStegFormer, a robust steganography framework based on an autoencoder architecture. It integrates a Cross-Attention Guided Global Enhance Bottleneck (CrossGEB) to improve feature fusion between cover and secret images. A stage-wise progressive perturbation strategy is used during training to gradually increase distortion difficulty, enhancing robustness. Additionally, a lightweight joint loss combining perceptual and statistical constraints improves visual quality and semantic consistency. Extensive experiments show that RobustStegFormer achieves strong performance under multiple distortions. Specifically, it improves secret image PSNR by more than 2.5 dB under Gaussian noise (<tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\sigma=10$</tex>) and over 4 dB under JPEG compression (<tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\text{QF}=80$</tex>), outperforming state-of-the-art models and demonstrating its practical effectiveness.