SigLDiff: A Signature Based Latent Diffusion Model for Forged-Free Offline Signature Verification

Mingjian Zhang, Lidong Zheng, Jiaen Chen, Yichi Zhang, Yuchen Zheng

Abstract: In real Offline Handwritten Signature Verification (OHSV) scenarios, skilled forged signatures are extremely rare and cannot be accessed during model training and verification stages. Therefore, the difference between genuine signatures and skilled forgeries is hard to capture in this scenario. In this paper, a forged signature generation method named SigLDiff based on latent diffusion models is proposed to address the challenge that skilled forgery cannot be accessed in OHSV tasks. SigLDiff constructs a forged signature generation approach with fine-grained variations by disentangling the content features of genuine signatures from the style characteristics of other genuine signatures. Specifically, the proposed method initially employs the multiscale content aggregation module to integrate the global and local stroke feature of genuine signatures, which effectively enhances the detailed information of signature strokes at different scales. Moreover, to capture style features of signatures, a cross-scale style representation module is proposed to extract handwriting style features at various abstraction levels via a pre-trained VGG network. As a result, style transfer of genuine signatures is implemented in the latent space, enabling generated samples to preserve the content features of the target user while exhibiting diverse forgery styles. Extensive experimental results on five public datasets demonstrate that verification models trained with our generated samples defeat models without authentic skilled forgeries.

External IDs:dblp:conf/icdar/ZhangZCZZ25