Lip-based biometric authentication

Abstract

Lip movements offer a wide range of benefits as a potential biometric for authentication, especially considering the emerging advances in mobile technologies and applications containing sensitive information. Lip movements can be captured using the front facing camera on mobile devices or laptops, requiring no specialist hardware and can be used in combination or as a sole biometric. Lip movements are a behavioural biometric which makes it hard for imposters to mimic or replicate and liveness detection can be easily added to any lip-based authentication solution making it more secure than it’s physiological-based contenders.

Despite this lip-based authentication is not a well researched topic. It has seen little work in comparison to other biometrics such as face, fingerprint or voice. A review of the literature suggests that the field has been held back by inconsistent results and lack of available data, particularly real-world data. There are no robust benchmarks to compare results. This has made it increasingly difficult to tell where progress in the area has been made.

To try and address these issues this work investigates visual information from video input of the lips as a biometric for authentication. This work confirms that there is uniqueness that can be taken from the lips for biometric authentication, and sets a robust benchmark with a popular and widely available dataset using a Gaussian Mixture Model-Universal Background Model approach. New data was captured for the work in this thesis and used to investigate lip-based authentication under challenging real- world conditions. The results were analysed to provide a deeper understanding of the behaviours and shortcomings. This work goes on to push the field forward through the application of deep learning. A deep artificial neural network is trained end-to-end, and one-shot-learning applied by implementing a Siamese network architecture. This approach achieved state-of-the-art performance for lip-based biometric authentication on the XM2VTS dataset.

Date of Award	Dec 2019
Original language	English
Awarding Institution	Queen's University Belfast
Supervisor	Darryl Stewart (Supervisor)