GENERATIVE AI FOR SIGNATURE SPOOFING

Abstract

Biometric data reveals characteristics of human traits that are helpful for identifica tion and verification purposes. A handwritten signature is a type of biometric data that is widely used for online and offline verification purposes. Tasks utilizing this data have been automated for various applications using Machine and Deep learning. Deep learning systems learn to recognise a person’s original signature and forged signature to ultimately be able to classify an image as an original or forged signature. To make these systems stronger, more data is needed and better quality data is needed, i.e., the forgeries have to be skilled enough so they are tougher to classify. This made researchers turn to generative techniques to enable forgery generation to increase the amount of data used for verification tasks and also make higher-quality forgeries in the process. Among other generative techniques, Generative adversarial networks (GANs) learn from gen uine and forged signatures to generate forged signatures. This technique creates a strong signature verifier which is basically the discriminator model of the GAN. However, work in this area neglects the fact that a forgery can not be too similar or dissimilar to the actual signature because that risks being rejected by real-life verification systems. Our research focuses on creating a generator that produces such forged samples that achieve a benchmark in spoofing signature verification systems. We use CycleGANs infused with Inception model-like blocks with attention heads and the SigCNN model as a base Discriminator. We train our model with a new paradigm shifting technique. We evaluate the “goodness” of our forgeries by creating a generic evaluation metric that utilizes influential points of the distributions of original samples and forged target sam ples to determine the closeness of generated forged images to both data domains. Our model successfully surpasses state-of-the-art image generation architectures in spoofing signature verification systems.