A customized CNN model for signature authentication—Forensic implications

Signature authentication refers to the verification whether the signature is legitimate or forged and is executed by the person who is claiming to have signed it. The main objective of the present research was to customize a deep learning-based convolutional neural network (CNN) model trained on a primary dataset for signature authentication. The model was trained, validated, and tested on the dataset of the 1400 signature images (700 genuine and 700 forged) primarily obtained from the study participants. The signature samples were equally divided into a train (1000 samples comprising 500 forged and 500 genuine signatures) and a test dataset (400 samples comprising 200 forged and 200 genuine signatures) as per the requirements of the model's training and testing procedure. The model's architecture was optimized with different hyperparameters to achieve a higher accuracy rate. The results show that the proposed model attains the training, validation and testing accuracy of 97.32%, 97.92%, and 84.5% respectively. Furthermore, other evaluation matrices were also computed, including precision, recall (sensitivity), F1-score, and specificity with the values of 85%, 84%, 84%, and 90%, respectively. The accuracy matrices provide better performance over the other existing methods. This customized CNN architecture can be trained on larger datasets and directly deployed in the field of forensic science for signature examination. The study has wide-ranging applications in the banking sector, forensic document examination, courtrooms, and beyond.