A signature is an important biometric datum used to verify identity in a range of areas, including banking and law. With the availability of solutions that enable machine replication of signatures, the need to reliably distinguish a genuine signature from a machine-made one is growing. The aim of this thesis is research and development of a software solution using artificial intelligence models to detect machine-generated signatures.
For this purpose, a unique dataset was created containing RGB images as well as 3D structures of signatures obtained using a specialized forensic device. The combination of these modalities allows removal of noise and unwanted artifacts. Together with the application of data augmentation, this enabled training of more robust models.
This thesis presents several neural network methods based primarily on convolutional layers and clustering methods. The paper implements xAI techniques using GradCAM and SHAP methods. By comparing the results of the methods, the following were selected: a fine-tuned Siamese convolutional neural network trained on a Czech dataset with 100 % accuracy, a fine-tuned VGG-16 model with 95.8 % accuracy, shallow convolutional neural networks with 87.5 % accuracy, and a random forest classifier with 75 % accuracy. The results may be influenced by the size and variability of the test set, which can lead to an overestimation of the model’s ability to generalize learned patterns and over-fitting during training phase, as well as by the selection of manually extracted features.