A Robust Approach for Handwritten Signature Verification Through Deep Learning

More current research shows a rise in the use of digital signatures to deceive, therefore an improvement in forgery incidences. This research proposal seeks to establish and optimize efficient and accurate deep learning methods for distinguishing between real and forged handwritten signatures in order to improve security in online transactions. Three approaches are explored: MobileNet that is type of Convolutional Neural Network (CNN) and combined architecture of MobileNet with LSTM. MobileNet is selected because of its small size and effectiveness in extracting features from signature images. The CNN model used in this work involves stacking several layers of convolution in order to abstract unique and fine features of handwritten signatures. The sequences increases the temporal behavior knowledge between the features that improves the forgery exposure based on the relations of components of the feature matrices.The dataset used for the purpose is a very pragmatic and constitutes of actual signatures along with forged signatures which provides full fledge training along with efficient testing. This dataset encompasses 2200 authentic and forged signatures, which offers a solid basis for both training and testing. Some of the procedures applied are data normalization and data augmentation in their preparatory applications, increasing the performance of the model. The performance analysis, done in terms of accuracy, precision, recall, and F1-score, indicates that the CNN model performs better than either MobileNet or hybrid models. In summary, training accuracy of 99.77% and a training loss of 0.0115 for the CNN model, 98.39% as the accuracy with 0.0582 training loss for the MobileNet model, and the hybrid model with 77.19% training accuracy and 0.4876 training loss. All of this recognizes a valid means of making this hybrid approach effective in lowering the risk of forgery and improving digital signature security.