An Efficient mobile handwritten signature verification system based on Convolution Neural Netwoks (CNN)

Abstract

Signatures are used as the main form of identification of the owner of the activity that is performing a transaction. A signature can only be accepted if it is from the authorized person. It is highly unlikely that two signatures created by the same person are identical. Many signature properties may vary even when the same individual signs two documents. As a result, detecting a forgery becomes a challenging task. Handwritten signature verification faces various challenges, including signature variability and the risk of forgery. Fraud through forgery is a common issue that drives corporates and businesses into significant financial loses and affects brand reputation in various sectors such as banking and government agencies which deal with important physical and online signed documents, legal paper works and government policies on daily basis Therefore, there is a need for a robust and accurate handwritten signature verification system that achieves the highest degree of accuracy in detecting fraud through forgery. The main objective of the research is to design, develop, test and validate the performance of an Efficient Mobile Handwritten Signature Verification System Based Convolution Neural Network, which can detect the authenticity of handwritten signatures with accuracy, precision and in real-time. The study is underpinned by theories such as Machine learning and Deep Learning Theory and Risk Adaptation Theory (RAT). The study utilized an Iterative Development Approach combined with Object-Oriented System Analysis and Design (OOSAD). The methodology involved incremental cycles of planning, analysis, design, implementation, testing, and evaluation. Convolutional Neural Networks (CNNs) was used for signature verification, utilizing preprocessing techniques like grayscale conversion, binarization, and noise removal. Signature datasets were sourced from Kaggle and AI-generated sources. The frontend was developed using Flutter, while the backend leverages Laravel API for secure communication. The model was trained and validated using performance metrics such as accuracy, precision, recall, and F1-score. Testing included system validation, user acceptance testing (UAT), and security checks. The study found that the CNN-based model coupled with a flutter-based mobile application achieved a 97.2% accuracy in distinguishing genuine from forged signatures, outperforming traditional methods. The system provides real-time verification, making it suitable for financial and legal applications. Additionally, strong security measures like encrypted data transmission and user authentication prevent unauthorized access. Keywords: Binarization; Image preprocessing; F1-score; precision; Recall; Forgery; fraud; signature verification; Mobile system