Title of the Talk: Closing the Gap: The Integration of Synthetic and Real-World Facial Data
Speaker: Brian Lovell, University of Queensland, Australia
Abstract: New research indicates that datasets trained solely on synthetic data can reach performance levels within a mere 3% of those trained on real data. Our recent findings go a step further, demonstrating that purely synthetically trained datasets can surpass this threshold, suggesting a future where synthetic data becomes the standard for training. This breakthrough liberates face recognition from its dependence on traditional photography, paving the way for synthetic systems that could outperform their photographic counterparts. Synthetic face systems offer unparalleled flexibility in manipulating factors often challenging to alter in reality. For instance, they allow for the creation of faces sporting diverse hair colors or facial hair styles, a feat impractical with real individuals on a large scale. Such capabilities hold the potential for accuracy enhancements beyond what traditional photographic methods can deliver.
Title of the Talk: Biometric Sample Quality: the Recognition Model Perspective
Speaker: Naser Damer, Fraunhofer Institute for Computer Graphics Research IGD, Germany
Abstract: Biometric sample quality measures the utility of the sample to the recognition algorithm. The relevance of the sample quality to different processes in biometric systems attracted the attention of different stakeholders. This interest was clear in the recent NIST evaluation “Face Analysis Technology Evaluation (FATE) Quality” and the ever-evolving ISO standardization effort, including the ISO/IEC 29794-1. Given face recognition as an example, this talk will first present an overview of the main aspect of modern face recognition models, their training, and the template inference process. Based on that, the talk will discuss different aspect of the recognition model behavior that can present indications related to sample quality. Such behavior observations are perceived both in the inference and training process of modern face recognition models and, if probably understood, can be translated into quality scores inherently indicating the utility of biometric samples.
Title of the Talk: Recent Advances in Privacy-Preserving Biometrics Authentication
Speaker: Jiankun Hu, University of New South Wales, Australia
Abstract: It is well-known that biometrics can provide an authentication of genuine users. Significant advances have been made in the field with many successful applications, e.g., border control and digital access control. Biometrics involves a person’s privacy data which is regulated by laws in many countries. There is a trend/need to develop privacy-preserving biometrics authentication technologies. In this talk, I’ll introduce some major research works in this field. It will cover the popular infinite-to-one mapping-based cancelable biometrics template design, Attack via Record Multiplicity (ARM), ARM attack resilient cancelable biometrics designs, and hill-climbing attacks on biometrics templates. We will introduce our latest projects/works on ARM and hill-climbing resilient cancelable deep learning models, and Biometrics-Based Authenticated Key Exchange Protocol with Multi-Factor Fuzzy Extractor (if time permits).