SignaShield: Guarding Your Signature via Font Style Watermarking
Rensong Wang, Jie Zhang, Zhiwen Ren, Weiming Zhang, Nenghai Yu · 2025
Image-based signatures, which represent a person’s unique handwriting or writing style in image form, have become increasingly common in modern digital transactions. These signatures are widely used for authentication and authorization, offering convenience and efficiency across various domains. However, their growing prevalence also introduces significant security risks. Attackers may exploit these image-based signatures to forge identities, commit financial fraud, or gain unauthorized access. The risks are further exacerbated by the advancement of AI models, which leverage data-driven techniques to replicate handwriting styles with remarkable accuracy, making such attacks increasingly feasible. To address it, watermarking presents a straightforward and practical solution. However, existing methods typically embed watermarks across the entire image, making them ineffective against AI models, which can isolate and replicate the style without preserving the watermark.In this paper, we propose SignaShield, the first watermarking method specifically designed to protect image-based signatures by embedding watermarks directly into style features. Our approach integrates a watermarking module into existing style transfer networks, leveraging their generative capabilities to produce high-quality, watermarked signature images. This design enables the recovery of the original watermark even when attackers attempt to imitate styles using watermarked images. Additionally, we identify and exploit redundancy in the latent space of style transfer networks, allowing for seamless and inconspicuous watermark embedding. Extensive experiments and ablation studies validate the effectiveness of SignaShield, demonstrating superior synthesis quality, robust watermark extraction, and strong style preservation.