Offline handwritten signature authentication using Graph Neural Network methods

Due to their uniqueness and simplicity, handwritten signatures are used as a behavioral biometric feature to identify and authenticate individuals. Due to the increase in the criminal activity of forgers in forging signatures, organizations are forced to use computer systems to verify the authenticity of signatures. For this reason, offline signature verification system is widely used in most organizations. Despite the abundance of research conducted on signature verification, it is difficult to distinguish real and forged signature samples due to the lack of information in the signing process. On the other hand, the small number of training samples is a challenge for offline signature recognition systems. In recent years, to improve these problems, systems based on machine learning and deep learning methods have been presented. In this paper, we have proposed a graph neural network-based architecture for offline signature verification. In this work, the features in the signature images, which are the pixels that make up the signature, are extracted by the SIFT algorithm and sent to the graph-based neural network as a graph structure. After training the network, the data of the test samples are classified into one of two classes, genuine or forged. The proposed model was evaluated on two datasets, MCYT-75 and UTSig, and Accuracy (Acc), Average Error Rate (AER), False Acceptance Rate (FAR) and False Positive Rate (FPR) were considered as performance measures. In this model, the values of Acc, AER, FAR and FPR for the MCYT-75 data set are equal to 1,0, 0, and 0, respectively, and for the UTSig database, these values are equal to 0.092, 0.007, 0.014 and 0.