A Modified Swin Transformer With Pairwise Attention for Offline Signature Verification

Offline signature verification (OSV) remains a challenging biometric task, particularly under high intra-writer variability and skilled forgery conditions. Most deep learning–based approaches encode reference and query signatures independently, which may limit their ability to model fine-grained correspondences. This paper presents a modified Swin Transformer architecture that introduces an encoder–decoder hierarchy with multi-scale feature aggregation and a pairwise attention mechanism to enable explicit reference–query interaction. The proposed model is trained using a focal contrastive loss under a strict writer-independent protocol. Experiments conducted on the UTSig and CEDAR datasets evaluate performance using accuracy, false acceptance rate, false rejection rate, and equal error rate. Results show consistent reductions in false acceptance rates and modest improvements in overall verification accuracy relative to a Swin Transformer baseline, accompanied by changes in false rejection behaviour. Additional analyses, including learning dynamics, confusion matrices, and threshold sweeps, are used to examine the resulting security–usability trade-offs. These findings suggest that multi-scale cross-attention can contribute to improved discrimination in OSV under challenging variability conditions, while highlighting the importance of careful threshold selection in practical deployments.