Abstract
<title>Abstract</title> <p>Document image steganography is a key information-hiding task aimed at securely and effectively embedding secret information into document images while preserving visual quality. State-of-the-art document image steganography methods typically focus on feature fusion. However, they often overlook edge information when designing fusion strategies. In this work, we observe that the incorporation of edge features leads to improved fusion performance. Based on this observation, we propose a novel method: Edge-Guided Document Image Steganography (EGDIS). The proposed model first applies an adaptive Canny algorithm to extract high-quality edge maps from document images. Subsequently, a Transformer-based Cross-Attention Fusion Network (CAFN) is employed to perform multi-scale feature fusion among the raw image, edge image, and secret message. By leveraging cross-attention mechanisms, the model enables fine-grained integration across modalities and scales. This mechanism captures long-range dependencies in image content, facilitating the deep integration of secret messages with raw and edge images, while preserving edge distribution patterns and global semantic consistency. Extensive experiments demonstrate that EGDIS outperforms existing methods on both DocImgCN and DocImgEN datasets.</p>