Fake Signature Recognition Using Convolutional Neural Network with ResNet-18 Architecture

Abdul Aziz Al Basyir
Kenan Almas
Big Daya Yudha Asmara

Frontier Advances in Applied Science and Engineering
Vol. 3 No. 2 (2025)

Signatures remain one of the most widely accepted forms of personal authentication in official document validation. However, offline signature recognition is inherently challenging because genuine signatures may vary across instances, while forged signatures can closely resemble authentic ones. These conditions make manual verification subjective, inefficient, and susceptible to human error, especially in high-volume document processing. This study proposes an automatic offline signature recognition approach using a deep learning framework based on a Convolutional Neural Network (CNN) with a ResNet-18 architecture. A total of 1,038 signature images were used and divided into training, validation, and testing sets to evaluate model performance under different training epochs and optimization settings. The experimental results show that the best performance was achieved by the Adam optimizer at epoch 25, with an accuracy of 59.5%. Nevertheless, the overall classification performance remained limited, indicating that the model struggled to learn sufficiently robust discriminative features for reliably separating genuine and forged signatures. These findings suggest that offline signature recognition remains a difficult task when formulated as a simple binary image classification problem, particularly under conditions of high visual similarity between classes and limited data variation. The study highlights the need for improved training strategies, more suitable verification-oriented protocols, and more comprehensive evaluation metrics to enhance the reliability of automatic signature recognition systems.