Publication: Digitally captured signatures verification using transfer learning with different types of capacitive stylus pens
Loading...
Date
2022-08-01
Authors
Zaidi, Nur Auni Anisah
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Over the last few years, there has been an increasing interest in using personal identity such as signatures as an authentication method. Since signatures can often be forged by others, the signature verification system is developed to determine whether the signature is genuine or forgery. Digitally captured signatures (DCS) are signatures made on electronic devices such as smartphones and tablets, which the dynamic features of the signatures are insufficient making it more challenging compared to online signatures. In this research, a DCS verification using transfer learning with different types of capacitive stylus pens is proposed. The pre-trained transfer learning model is fine-tuned to suit the task of DCS verification. The performance parameters used to evaluate the performance of DCS verification are Accuracy (ACC), False Acceptance Rate (FAR), False Rejection Rate (FRR), Average Error Rate (AER), and Equal Error Rate (EER). In addition, new DCS datasets are made. The signatures will be gathered from 30 individuals with each one of them will make 30 genuine signatures. Then, another 30 individuals will make 30 forgery signatures based on the genuine signatures made. All signatures made will be using the Baseus and Goojodoq capacitive stylus pens, respectively. The fine-tuned GoogLeNet achieved better performance with 95.17% ACC, 7.67% FAR, 2.00% FRR, 4.83% AER, and 0% EER
when tested on and compared to Tam (2021) DCS dataset. The DCS verification performance is then compared by testing Convolutional Neural Network (CNN) architecture and GoogLeNet on new DCS datasets and the result shows that the GoogLeNet achieved better accuracy of 95.83%, 95.17% and 95.92% for Baseus, Goojodoq and mix of both pens datasets. Furthermore, the proposed GoogLeNet is tested on CEDAR dataset, achieved better accuracy of 98.46%, 97.96% and 99.20% for when the training:testing ratio and maximum epochs are differed. The performance is compared with state-of-art VGG-19 and CNN architecture.