Abstracts

In recent years we have witnessed increasingly diverse application scenarios of biometrics systems in our daily life, one of which is person recognition at a distance. In this talk, I will cover a number of key problems that are recently being addressed at the Computer Vision Lab of MSU, including: how to address the various training sample quality in learning large-scale face recognition systems; how to integrate identity information from an image set or video sequences; how to estimate the 3D body shape from an image of clothed human body; how to use AIGC to generate a complete synthetic database for training face recognition systems; how to leverage foundation models for person recognition; and how to build our own foundation models for unified face and body recognition.

Hollywood's favorite biometric attack often involves an eye plucked from its socket and presented on a stick to an iris camera. But... can this attack actually succeed in reality? Can we use iris patterns to recognize humans after death? And if so, how long after death is this possible? Looking at the other end of life: can we apply iris recognition for newborns? Can we learn from human experts to develop iris recognition and presentation attack detection methods that generalize well to unseen data? Can we synthesize realistically-looking irises of non-existent identities and condition these models with specific anomalies to assist forensic experts in their work? Join us on a biometric journey that starts at birth and ends after death. Along the way, we'll try to find answers to the above questions, supported by the latest research advancements in forensic iris recognition.

As biometric applications become more prevalent in our everyday lives (e.g., border crossing, unlocking of smartphones, banking), it is important to consider how our irreplaceable biometric data should be protected when it is used (stored and processed) in these biometric systems.  Traditional data protection mechanisms, such as cryptographic hashing and encryption, are unsuitable due to their sensitivity to small changes in the input data -- a property that conflicts with the "fuzzy" nature of biometric measurements.  For this reason, the field of Biometric Template Protection (BTP) was conceived over two decades ago, with the aim of developing protection techniques that are more suitable to the nature of biometric data.  In this presentation, I will present an introduction to this field of research.  In particular, I will aim to answer the following three questions:

1.) What is Biometric Template Protection (BTP)?

2.) Why do we need BTP?

3.) How to achieve BTP?

After a brief introduction on the overall architecture of a fingerprint recognition system, this lecture presents the classical fingerprint feature extraction and matching steps: segmentation of the ridge-line pattern from the background, estimation of local ridge-line orientation and frequency, enhancement of the ridge-line pattern, minutiae detection and minutiae matching. New (deep learning based) approaches are then introduced, and the pros/cons of classical vs deep-learning approaches are discussed. 

I will present a wide variety of work on the topic of Face image generation, including diffusion models and GANs.

H. Chefer, O. Lang, M. Geva, V. Polosukhin, A. Shocher, M. Irani, I Mosseri, L. Wolf. The Hidden Language of Diffusion Models. International Conference on Learning Representations (ICLR), 2024.

Y. Tewel, O. Kaduri, R. Gal, Y. Kasten, L. Wolf, G. Chechik, Y. Atzmon Training-Free Consistent Text-to-Image Generation. ACM SIGGRAPH, 2024.

E. Sheffi, M. Rotman, L. Wolf Gradient Adjusting Networks for Domain Inversion. Scandinavian Conference on Image Analysis (SCIA), 2023. .

R. Shmelkin, T. Friedlander, L. Wolf. Generating Master Faces for Dictionary Attacks with a Network Assisted Latent Space Evolution. IEEE International Conference on Automatic Face and Gesture Recognition (FG), 2021.

and maybe a deep fake detection work

Y. Nirkin, L. Wolf, Y. Keller, T. Hassner. DeepFake Detection Based on Discrepancies Between Faces and their Context. IEEE Transactions on Pattern Analysis and Machine Intelligence (PAMI), 2021. 

Foundation models have revolutionized AI by enabling broad generalization across tasks and modalities. In this talk, I will discuss recent advances in applying foundation models to biometrics. I will highlight how large-scale pretraining, transfer learning, and multimodal integration are shaping the future of biometric systems.

In biometrics, Presentation Attacks (PA also referred to as spoofing) are performed by falsifying the biometric trait and then presenting this falsified information to the biometric system, one such example is to fool a fingerprint system by copying the fingerprint of another person and creating an artificial or gummy finger which can then be presented to the biometric system to falsely gain access. This is an issue that needs to be addressed  because it has recently been shown that conventional biometric techniques are vulnerable to presentation attacks. One of the main challenges in Presentation Attack Detection (PAD also referred to as anti-spoofing) is to find a set of features and models (mostly classifiers) that allows systems to effectively distinguish signals that were directly emitted by a human from those reproduced by an attacker. This talk will present an overview of typical face PAs and PAD techniques.

With the rapid progress of recent years, techniques that generate and manipulate multimedia content can now provide a very advanced level of realism. Powered by large language models, text-driven synthesis tools allow the user to modify or create from scratch images and videos by means of simple text instructions. On the one hand, this opens the door to a series of exciting applications in different fields such as creative arts, advertising, film production, video games. On the other hand, it poses enormous security threats. Therefore, there is an urgent need for automated tools capable of detecting false multimedia content and avoiding the spread of dangerous false information. This talk aims to present an analysis of the methods for synthetic media verification. Special emphasis will be placed on the phenomenon of AI-generated content and on modern data-driven forensic methods to fight them. The analysis will help highlight the limits of current forensic tools, the most relevant issues, the upcoming challenges, and suggest future directions for research.

This talk first examines the issue of how face recognition accuracy is different across demographics. Then we consider the common assumption that demographic accuracy variation arises due to demographic imbalance in the training data. Finally, we evaluate the role of automated face recognition in high-profile instances of wrongful arrest. In each area, empirical evidence may not support some popular viewpoints.

The last decade has delivered orders of magnitude improvements to the accuracy of biometrics algorithms. These improvements have been fueled by advancements in deep learning architectures and hardware. While they have allowed systems to operate at greater scales, accuracy, and ease-of-use, these opportunities have created new and more complex challenges than were the focus of biometric technologies of the past. In this talk we will first discuss the performance of state of the art algorithms. From there we will focus on the various challenges for deploying these algorithms into operational use-cases. As the talk progresses we will flag several key areas that need substantial academic research focus in the years ahead.