# Improving the Security of Smartwatch Payment with Deep Learning

**Authors:** George Webber  
**arXiv:** [2307.05437](https://arxiv.org/abs/2307.05437)  
**Date:** 2023-07-11  
**Summary page:** https://www.summarizepaper.com/en/paper/2307.05437v1/

## Key points

- Growing popularity of contactless payments using smartwatches
- Lack of traditional biometric security measures in smartwatch payments
- WatchAuth proposed as a system to authenticate smartwatch payments based on physical gestures
- WatchAuth requires a lengthy enrollment process for acceptable error levels
- George Webber's dissertation explores the use of deep learning to reduce the number of gestures required for enrolling into an authentication system for smartwatch payments
- Construction of a deep-learned authentication system that surpasses current methods with limited gestures provided by the user
- Development of a regularized autoencoder model that generates synthetic user-specific gestures
- Incorporating synthetic gestures improves the classification ability of the authentication system
- Technique allows for reducing the number of gestures needed to enroll a user into a WatchAuth-like system without negatively impacting error rates
- Research aims to enhance security and convenience in smartwatch payments through optimized gesture-based authentication systems and significantly reduced enrollment time while maintaining accuracy levels.

## Summary

In the field of contactless payments using smartwatches, there is a growing popularity but a lack of traditional biometric security measures such as facial or fingerprint recognition. To address this issue, Sturgess et al. proposed WatchAuth in 2022, a system that authenticates smartwatch payments based on the physical gesture of reaching towards a payment terminal. However, this system requires users to undergo a lengthy enrollment process to achieve acceptable error levels. In this dissertation by George Webber, the author explores whether deep learning can be applied to reduce the number of gestures required for enrolling into an authentication system for smartwatch payments. The first step is constructing a deep-learned authentication system that surpasses the current state-of-the-art methods, even when the user has provided only a limited number of gestures. Additionally, the author develops a regularized autoencoder model that generates synthetic user-specific gestures. By incorporating these synthetic gestures into the training process, it is demonstrated that the classification ability of the authentication system improves. This technique allows for reducing the number of gestures needed to enroll a user into a WatchAuth-like system without negatively impacting its error rates. Overall, this research aims to enhance both security and convenience of smartwatch payments by leveraging deep learning techniques to optimize gesture-based authentication systems and reduce enrollment time significantly while maintaining accuracy levels at an acceptable rate.