Abstracts

"Foundation models" are large-scale machine learning models - typically based on deep neural networks - that are trained on vast amounts of data that can be adapted to a wide range of problems. The term gained prominence with the rise of neural network models like GPT, BERT, and CLIP, which serve as the base or "foundation" for many different applications involving text, image, audio and video. Foundation models have transformed the field of AI. In this talk we will discuss how these models are being used in biometrics, which involves recognizing humans based on their biological and behavioral traits such as face, fingerprint, iris, voice, and gait. In particular, we will consider biometric tasks such as segmentation (e.g., iris localization), attribute extraction (e.g., age, sex, race, health from face images), and matching (e.g., face or iris recognition). Finally, we will enumerate the advantages and disadvantages of utilizing foundation models in the context of biometrics.

• Deep Learning in the Field of Biometric Template Protection - Prof. Christian Rathgeb:

This presentation reviews how deep learning reshapes biometric template protection. It introduces the fundamentals of biometric systems and highlights the need for safeguarding sensitive biometric data beyond conventional cryptography. Core template protection methods, biometric cryptosystems, cancelable biometrics, and encrypted-domain approaches, are outlined. The presentation discusses how deep neural networks enhance biometric performance and privacy by improving feature extraction and enabling feature type transformations such as quantization, binarization, and deep hashing. In addition, recent deep learning-based attacks, including reconstruction and “master sample” generation, are presented alongside countermeasures like deep pre-alignment and data minimization. Overall, the talk underscores both the benefits and risks of deep learning for biometric systems, stressing

• Biometric Template Protection using Homomorphic Encryption - Prof. Adi Akavia
  
  Homomorphic Encryption (HE) enables computations to be performed directly on encrypted data, thus preserving confidentiality throughout processing. This property makes HE particularly attractive for protecting biometric data—not only during storage and transmission, but also during matching and identification operations. However, the high computational and storage costs associated with HE remain major practical challenges.

In this talk, we will review recent advances that mitigate these overheads and discuss their potential implications for scalable privacy-preserving biometric identification systems.

• Recovering Facial Templates from Observing the Binary Result of Authentication Attempts - Prof. Orr Dunkelman
  
  

Biometric data is considered to be very private and highly sensitive. As such, many methods for biometric template protection were considered over the years --- from biohashing and specially crafted feature extraction procedures, to the use of cryptographic solutions such as Fuzzy Commitments or the use of Fully Homomorphic Encryption (FHE).

A key question that arises is how much protection these solutions can offer when the adversary can inject samples and observe the outputs of the system. While for systems that return the similarity score, one can use attacks such as hill-climbing, for systems where the adversary can only learn whether the authentication attempt was successful, this question remained open.

In this work, we show how to recover the facial templates from such binary results. We show that one can reconstruct a template with a very small reconstruction loss.This suggests that even when using advanced protection mechanisms, such as FHE, attribute-based encryption, MPC, or fuzzy extractors, the adversary is capable of reconstructing a facial image. Hence, the reliance only on cryptographic mechanisms is highly discouraged.
This is a joint work with Margarita Osadchy and Eliron Rahimi

• Face Image Quality Assessment - Prof. Christoph Busch:

The talk will address challenges of face recognition systems. When dealing with operational systems, the quality of captured face images is relevant as it will impact the recognition accuracy. Thus, it is required to measure the utility of a face sample  with a quality score but also with complementary measures that can provide actionable feedback. The ISO/IEC 29794-5 defines a unified quality score (UQS) for a face image and numerous, component quality measures (CQM), which provide accurate information such as the correctness of the pose or the sharpness of the face image (and many others). The open source software OFIQ as reference implementation of ISO/IEC 39794-5 can validate, whether a face image is compliant to ICAO requirements.

• When Faces Lie: The cat-and-mouse game between face forgeries and forgery detection - Dr.-Ing. Clemens Seibold
  
  We trust faces to recognize each other in everyday life and as a key biometric trait for identity verification. Advances in computer graphics and generative AI are increasingly breaking this trust, as they enable the creation of highly realistic forged faces in images and videos. These forgeries challenge the reliability of visual media in society and the trustworthiness of facial recognition in security contexts.

This talk provides an overview of detection methods with a strong focus on explainability for such forged media. It outlines current techniques for detecting synthetic faces and offers insights into recent trends in computer graphics and AI for synthesizing and animating realistic 3D heads that are becoming ever more convincing and difficult to detect.

•  Self-ID: Authentication Based on Visual Self-Recognition - Dr. Hendrik Graupner
  
  

Authenticating users securely and intuitively remains a central challenge in bridging physical and digital identities. Recent advances in cognitive biometrics suggest that individual visual face recognition can serve as a distinctive and measurable biometric trait. Building on our pre-study conducted in 2022 that confirmed the feasibility of measuring visual face recognition through eye tracking, a new research project funded by the German Cyberagentur was launched in May 2025. This project investigates face- and self-recognition–based cognitive traits as novel biometric modalities, aiming to translate these findings into practical, privacy-preserving authentication solutions. Beyond conventional authentication, one particularly promising application is the detection of deepfakes in video-conferencing environments, where visual self-recognition mechanisms could help distinguish real participants from manipulated video streams. The talk will outline the scientific foundations, preliminary findings, and envisioned use cases of this emerging direction in biometric cybersecurity.