Prototypical signatures improve offline signature verification, ICPR 2026
New method uses compact signature summaries to generate diverse training samples for forgery detection.
A team led by Kecia G. de Moura (with Robert Sabourin and Rafael M. O. Cruz) has developed a new method for offline handwritten signature verification, accepted for oral presentation at ICPR 2026. The work addresses a key challenge: real forgeries are rarely available, so negative samples are typically drawn randomly from other users' genuine signatures. This random selection often lacks diversity and increases redundancy, making training inefficient.
The researchers introduce prototypical signatures—compact, non-identifiable summaries of genuine signature features—to generate diverse and informative negative samples. Their experiments show three key findings: (i) these prototypical signatures yield more informative negative samples, significantly improving detection of skilled forgeries; (ii) the method is backbone-agnostic, meaning it works robustly across various neural network architectures; and (iii) when combined with a primal-form linear SVM, it becomes a compelling alternative to RBF-based models, offering major gains in scalability and computational efficiency. The code is available on GitHub.
- Prototypical signatures generate diverse negative samples, improving skilled forgery detection over random selection.
- Method is backbone-agnostic, functioning across different neural network architectures.
- Combined with linear SVM, it matches RBF-based performance with better scalability and lower computational cost.
Why It Matters
This approach could make automated signature verification more reliable and efficient, with applications in banking and legal document authentication.