Towards Open-world Face Presentation Attack Detection: An Asymmetric Bi-Meta Learning for Domain Generalization Approach

Project: Research project

Project Details


Background and Motivations: Face recognition technology has been extensively utilised for many practical applications in the past few years. For example, electronic payment using face recognition, which was selected by MIT Technology Review as one of the top-ten breakthrough technologies in 2017, has been fully deployed in mobile devices. However, whilst face recognition is very convenient for users, its security implications have not been fully explored. Specifically, it is unknown to what extent face-recognition systems are capable of identifying illegal users who present a fake face of an enrolled user. Such an attempt to fraudulently access a system is called a face presentation attack or a faces spoofing attack, and the technique to prevent it is denoted face presentation attack detection or face anti-spoofing. Research on face anti-spoofing technology has received less attention in the past decade, and most of the current research are focusing on detecting a single and known type of attack, such as image (print), video (replay) or 3D mask. Moreover, a detector in a real-world scenario has no prior information on the type of attack to be expected. Furthermore, publicly available fake-face datasets are limited and relatively small, typically comprising approximately hundreds to thousands of samples with tens to hundreds of individuals, which makes training anti-spoofing method difficult. As such, it is important to develop a method to detect a face-presentation attack that can function effectively with limited training data and in situations in which prior knowledge of the attack type is not available.

Problem Definition and Challenges: Based on the background and motivations given above, the open- world face anti-spoofing problem in this project is defined as follows. Given N types of labelled fake- real face-pair datasets with different within-type variations and many real face data for training, this project will develop a face presentation attack detector in which the type of attack is not known. The challenges addressed by this project will include: (i) an unknown attack type, which could be a combination of different types of attacks (e.g., a partial mask plus eyeglasses); (ii) the large within-type variations of the same type of attack (e.g., different mask materials for a 3D-mask attack); (iii) the limited number of fake-face training samples; and (iv) no labelled nor unlabelled face data from the test (target) domain for training.

Novelty of This Project: A domain generalisation approach will be adopted to solve these challenges. Since the number and size of publicly available fake-face datasets are small and limited, we need to enlarge the source domains by generating more fake-face attack type variations and combining different attack types. Generalized features can then be learned to handle unknown attack types. Inspired by the idea of learning to learn, a novel asymmetric bi-meta learning for domain generalisation method will be developed in this project. The proposed method will jointly optimise a set of enlarged fake-face source domains and discriminative generalisation features to afford a face presentation attack detector. This will involve the development of novel two-level asymmetric image- and feature-level meta-learners, and a novel fake-face data generation algorithm.

Long-term Significant: The results of this project will be timely and important contributions to the face biometrics research community and will potentiate the advancement of more practical facial biometrics security research, geared towards the generation of secure and reliable face-recognition systems. The new knowledge of meta-learning for domain generalisation that will be revealed during our proposed work will also contribute to the machine learning community. In addition, the proposed domain generalisation method using asymmetric bi-meta learning will see general use and will be suitable for use in other applications.
Effective start/end date1/01/2130/06/24

UN Sustainable Development Goals

In 2015, UN member states agreed to 17 global Sustainable Development Goals (SDGs) to end poverty, protect the planet and ensure prosperity for all. This project contributes towards the following SDG(s):

  • SDG 16 - Peace, Justice and Strong Institutions


Explore the research topics touched on by this project. These labels are generated based on the underlying awards/grants. Together they form a unique fingerprint.