Identifying Forged E-Signatures Using Convolutional Neural Network
John Daven R. Hernandez, Carlos Riel O. Mendoza, Noel B. Linsangan · 2024
Forgery detection is common nowadays and there is a need to recognize the authenticity of e-signatures. The study discusses a Convolutional Neural Network that is run on a Raspberry PI 3 and uses pre-trained RESNET 18 model to train on a small dataset of forged and authentic e-signatures to classify whether an input signature is forged or authentic. Data augmentation is used to increase the size of the input data. It creates variants of the sample signatures that can represent different ways the user writes their signature. Before the data is input to the model, contour detection is used on the image to make the features of the signature more prominent. Three models for different signatures are created and tested to determine the model's accuracy. After training, the model can classify signatures as forged or authentic. Confusion Matrix was used to obtain the results and the model collected an overall accuracy of 76.67 percent. It was noted that some incorrect authentic signature predictions can be caused by various inconsistencies with how the signature is written.