Using a Convolutional Neural Network to Detect Forged Signatures

Authors: Timothy Gribbin, Bin Li

SUNY Campus: Suffolk County Community College

Presentation Type: Poster

Location: UU 108

Presentation #: 85

Timeslot: Session C 1:45-2:45 PM

Abstract: Signature forgery detection plays a crucial role in authentication for financial transactions, legal documentation, and security-sensitive applications. In this research, we develop a deep learning-based model using convolutional neural networks to classify genuine and forged signatures with high accuracy. The dataset contains multiple samples of authentic and forged signatures from the same individuals (CEDAR dataset). This approach assumes that multiple known genuine signatures of an individual are available, enabling comparison-based detection. The model was trained on a labeled dataset using a supervised learning approach. Preprocessing steps included grayscale conversion and image normalization. We utilized a custom CNN architecture optimized for spatial feature learning, employing convolutional layers with ReLU activation, max pooling for dimensionality reduction, and fully connected layers for classification. The network was trained with 15 epochs using the Adam optimizer and binary cross-entropy as the loss function. The dataset was split into training and testing subsets, and the model achieved a test accuracy of 91%. While this approach enhances signature verification performance, its reliance on multiple reference signatures poses a challenge, extensive samples for every individual is not always feasible. This limitation makes it well-suited for applications where abundant signature records exist, such as sports memorabilia authentication. Future work aims to enhance the model to minimize the requirement for abundant signatures while maintaining high accuracy.