CausaOne-sign: Causal explainable one-shot signature verification with lightweight cross-modality fusion

Abstract

Background: Offline handwritten signature verification remains a difficult biometrics problem due to large intra-writer variability; skilled forgers; the limited number of reference samples available; and the black-box nature of many current deep learning based decision-making methodologies. Objective: To develop an interpretable, efficient one-shot learning framework that can perform offline signature verification for individuals who have never been seen before using as few reference signatures as possible. Materials and Methods: The proposed CausaOne-Sign model uses stroke aware graph encoding, transformer based reasoning, and prototypical embeddings, along with a causal attribution model to provide an explanation of how signature verification works. Experiments have been conducted using CEDAR, SigComp2011 UTSig, and BHSig260 datasets. Results: CausaOne-Sign achieved up to 97.4% accuracy and 99.1% area under the curve (AUC), with low ERR (1.8%), outperforming or matching state-of-the-art methods. Conclusion CausaOne-Sign offers a robust, interpretable, and resource-efficient solution for OSV, suitable for forensic and mobile applications.