Unmasking the imposters: towards improving the generalisation of deep learning methods for face presentation attack detection.

Identity theft has had a detrimental impact on the reliability of face recognition, which has been extensively employed in security applications. The most prevalent are presentation attacks. By using a photo, video, or mask of an authorized user, attackers can bypass face recognition systems. Fake presentation attacks are detected by the camera sensors of face recognition systems using face presentation attack detection. Presentation attacks can be detected using convolutional neural networks, commonly used in computer vision applications. An in-depth analysis of current deep learning methods is used in this research to examine various aspects of detecting face presentation attacks. A number of new techniques are implemented and evaluated in this study, including pre-trained models, manual feature extraction, and data aggregation. The thesis explores the effectiveness of various machine learning and deep learning models in improving detection performance by using publicly available datasets with different dataset partitions than those specified in the official dataset protocol. Furthermore, the research investigates how deep models and data aggregation can be used to detect face presentation attacks, as well as a novel approach that combines manual features with deep features in order to improve detection accuracy. Moreover, task-specific features are also extracted using pre-trained deep models to enhance the performance of detection and generalisation further. This problem is motivated by the need to achieve generalization against new and rapidly evolving attack variants. It is possible to extract identifiable features from presentation attack variants in order to detect them. However, new methods are needed to deal with emerging attacks and improve the generalization capability. This thesis examines the necessary measures to detect face presentation attacks in a more robust and generalised manner.

Name dc:type.qualificationname

PhD

Level dc:type.qualificationlevel

Doctoral

Grantor dc:publisher.institution

Robert Gordon University

Year dc:date.issued

2023

Author dc:creator

• Abdullakutty, Faseela Chakkalakkal

Advisor dc:contributor.advisor

• E. Elyan

• Facial recognition
• Presentation attacks
• Systems security
• Deep learning
• Neural networks

Language dc:language

en

Identifier

oai:rgu-repository.worktribe.com:2270640
https://doi.org/10.48526/rgu-wt-2270640

OAI identifier oai:identifier

oai:rgu-repository.worktribe.com:2270640