Handwritten Signature Verification Using Wrist-Worn Devices

This paper suggests a novel verification system for handwritten signatures. The proposed system is based on capturing motion signals from the sensors of wrist-worn devices, such as smartwatches and fitness trackers, during the signing process, to train a machine learning classifier to determine whether a given signature is genuine or forged. Our system can be used to: (1) Verify signatures written on paper documents, such as checks, credit card receipts and vote by mail ballots. Unlike existing systems for signature verification, our system obtains a high degree of accuracy, without requiring an ad hoc digital signing device. (2) Authenticate a user of a secure system based on "who you are" traits. Unlike existing "motion-based" authentication methods that commonly rely on long-term user behavior, writing a signature is a relatively short-term process. In order to evaluate our system, we collected 1,980 genuine and forged signature recordings from 66 different subjects, captured using a smartwatch device. Applying our signature verification system on the collected dataset, we show that it significantly outperforms two other state-of-the-art systems, obtaining an EER of 2.36% and an AUC of 98.52%.