Forged handwriting verification: a public domain dataset for training machine learning models

Handwriting verification is an important task in the legal framework where the paternity of handwritten wills or con-tracts needs to be ascertained. Due to the lack of automated tools, the court currently relies exclusively on handwriting experts, who often show a wide degree of uncertainty in their responses. Machine learning models, and in particular artificial Neural Networks (NNs) might be a valuable aid to experts by providing an objective and automated instrument for handwriting analysis, especially when expert witnesses are undecided. However, at the state of the art there is a scarcity of datasets which are suitable for training NNs for the handwriting verification task, preventing the development of models accurate enough to be introduced into forensic practice. In this paper, a dataset of 3320 genuine and 3320 forged handwritten samples from 166 subjects is presented, considering two scenarios: copy (imitation of handwriting maintaining the same text, 3320 samples) and spontaneous production (imitation of handwriting generating a new text, 3320 samples). We provide baseline results for deep siamese convolutional neural networks, that are deep learning models widely adopted in similar tasks. In the proposed dataset, such NNs were able to reach accuracies over 75% in identifying forged samples in the copy scenario and of almost 82% in the spontaneous production scenario. Finally, a sample of 550 humans were tested in the same classification task. The experiments show that NNs perform significantly better than humans in the spontaneous production scenario, which is more complex than the copy one. We publicly release our dataset to encourage the future development of advanced Deep Learning models for the forged handwriting verification task.