Telling Human and Machine Handwriting Apart

2025-10-01

IEEE Transactions on Systems, Man, and Cybernetics: Systems

AI Summary

This work proposes a shallow recurrent neural network–based approach to directly distinguish between human-written and machine-generated handwriting samples—including those produced by GANs, Transformers, and diffusion models—using raw pen-tip trajectories, thereby enabling authentic user verification and defense against automated attacks. It presents the first systematic evaluation of the discriminability of diverse handwriting synthesis methods in human–machine classification, eliminating the need for handcrafted features by integrating kinematic modeling grounded in the Sigma-lognormal framework. The method achieves an average AUC of 98.3% and an equal error rate of only 1.4% across ten public datasets. Remarkably, it maintains high performance on 90% of test sets using merely 10% of the training data and demonstrates strong robustness in cross-domain scenarios.

Technology Category

• Natural Language Processing: Code Generation / Program Synthesis from Natural Language
• Machine Learning: Learning with Manifolds
• Humans and AI: Human-in-the-loop Machine Learning

Application Category

• Responsible Web: Machine-in-the-loop, human agency and autonomy
• Economics, Online Markets and Human Computation: Data quality aspects of human-annotated datasets
• User Modeling, Personalization and Recommendation: Fairness-aware retrieval and ranking

Abstract

Handwriting movements can be leveraged as a unique form of behavioral biometrics, to verify whether a real user is operating a device or application. This task can be framed as a “reverse Turing test” in which a computer has to detect if an input instance has been generated by a human or artificially. To tackle this task, we study ten public datasets of handwritten symbols (isolated characters, digits, gestures, pointing traces, and signatures) that are artificially reproduced using seven different synthesizers, including, among others, the Kinematic Theory ( $\boldsymbol {\Sigma \Lambda }$ model), generative adversarial networks, Transformers, and Diffusion models. We train a shallow recurrent neural network that achieves excellent performance (98.3% Area Under the ROC Curve (AUC) score and 1.4% equal error rate on average across all synthesizers and datasets) using nonfeaturized trajectory data as input. In few-shot settings, we show that our classifier achieves such an excellent performance when trained on just 10% of the data, as evaluated on the remaining 90% of the data as a test set. We further challenge our classifier in out-of-domain settings, and observe very competitive results as well. Our work has implications for computerized systems that need to verify human presence, and adds an additional layer of security to keep attackers at bay.

Problem

• handwriting
• behavioral biometrics
• reverse Turing test
• human vs. machine
• synthetic handwriting

Innovation

• behavioral biometrics
• handwriting synthesis detection
• shallow recurrent neural network
• few-shot learning
• out-of-domain generalization