Enhancement of Siamese neural network for improved signature fraud detection

ABSTRACT: Offline signature verification presents persistent challenges in biometric authentication, hindered by degraded image quality, intra-class handwriting variations, and suboptimal decision boundaries that limit the reliability of automated systems. This study proposes and validates a systematically enhanced Siamese Neural Network (SNN) designed to overcome these limitations and improve signature fraud detection. The study’s approach integrates three targeted enhancements: (1) Contrast-Limited Adaptive Histogram Equalization (CLAHE) for processing, which significantly improves the visibility of fine stroke details in raw signature images; (2) Triplet Loss for representation learning, which constructs a more discriminative and well-structured embedding space by maximizing the margin between genuine and forged signatures; and (3) F1-score optimization for establishing a data-driven, adaptive decision threshold that balances precision and recall. The model was trained and evaluated on the publicly available CEDAR dataset using a writer-independent protocol. A modular evaluation demonstrated the individual and collective success of these enhancements. The final integrated model achieved an accuracy of 74.89%, a substantial improvement over the 51.49% accuracy of the baseline model. Importantly, the False Rejection Rate (FRR) was reduced from a prohibitive 91.21% to a practical 6.52%, signifying a major leap in system reliability and usability. These results confirm that the synergistic application of advanced preprocessing, representation learning, and thresholding techniques yields a more efficient and accurate solution for signature fraud detection, offering a scalable framework for secure biometric authentication in real-world applications.