Handwritten Signature Authentication Using Smartwatch Motion Sensors

The trade-off between security and ease of use tends to make passwords not necessarily as secure as designers expected. Biometric authentication has been receiving extensive attention and is increasingly used every day, of which signature authentication is one of the most commonly used methods due to the stability of signatures and the high difficulty of imitation. Current solutions often rely on dedicated digitizer consisting of graphic tablets and smartpens. The growth of commercial hand-worn devices such as smartwatches provides an alternative way to digitize signatures. Therefore, it is valuable to explore the feasibility of capturing the uniqueness and stability using hand-worn devices. In this paper, we propose a practical authentication method using smartwatch motion sensor data. It can distinguish whether an unknown signature belongs to the individual that they claimed to be or not. We firstly introduce Siamese Recurrent Neural Networks (RNNs) to deal with smartwatch motion sensor data of signing processes, which can save the task of manual feature design and improves system security. Our method uses a global model instead of a personalized one. Therefore, the trained system dose not require forged signatures from new users. After providing a set of genuine signatures during the enrollment phase, their signatures are irreversibly transformed into representation vectors, which will be used for authentication later while ensuring security. For experiment work, we collected 400 signature-related motion sensor data from 20 subjects and aligned them into 2990 pairs. Our method was evaluated using the collected data and outperformed comparable related work. We achieved an EER of 0.78%.