The speech will start by presenting eu-LISA’s core mission: the development and operation of EU’s large scale IT systems for identity management in the domains of Justice and Home Affairs (JHA).

Based on that foundation, the presentation will continue explaining how biometric recognition technology is inherently linked to eu-LISA’s main mission and present at the very core of all the systems managed by the Agency.

The presentation will then explore how, through biometrics, the interoperability of the systems managed by eu-LISA is expected to deliver several improved services and values to its users, in support of the dual goal of enhancing security while facilitating freedom of movement within the Schengen space and travel, focusing specifically on the functionalities of the shared Biometric Matching Service (sBMS).

The speech will then delve into the unique and difficult context in which eu-LISA operates in terms of business domains, database sizes, number of stakeholders, number of queries, very specific and strict regulatory framework.

Building upon the previous contextualisation, the presentation will lay out how, while much is expected and can be gained from the use of biometrics, the previously laid out unique context in which eu-LISA operates also brings several challenges that still need to be further addressed to provide a better service to the users and citizens, and to increase the added value that it brings to society. These challenges could shape current and future research in biometrics, and help bridge the gap between research under laboratory conditions and deployment in operational conditions. Among those challenges, some of the current priorities for the Agency are: 1) Biometric standarisation; 2) Biometric data quality; 3) Data protection;p 4) Algorithmic fairness; 5) Vulnerability protection.

The presentation will conclude with an open discussion on how those challenges are currently being addressed by research and academia to build the operational biometric systems of the future.

Despite growing attention to deepfake speech detection, the aspects of bias and fairness remain underexplored in the speech domain. To address this gap, we introduce the Speaker Characteristics Deepfake (SCDF) dataset: a novel, richly annotated resource enabling systematic evaluation of demographic biases in deepfake speech detection. SCDF contains over 237,000 utterances in a balanced representation of both male and female speakers spanning five languages and a wide age range. We evaluate several state-of-the-art detectors and show that speaker characteristics significantly influence detection performance, revealing disparities across sex, language, age, and synthesizer type. These findings highlight the need for bias-aware development and provide a foundation for building non-discriminatory deepfake detection systems aligned with ethical and regulatory standards.

Gait recognition is a biometric technology that identifies individuals based on their walking posture and motion. It has significant potential for applications such as criminal investigations and security systems. In silhouette-based gait recognition, silhouette normalization is a key component that enhances recognition accuracy. However, in scenarios where partial occlusions frequently occur, normalization can sometimes negatively affect recognition performance due to inappropriate adjustments. To address this issue, we propose a normalization refinement network based on spatial transformer networks, designed to mitigate the adverse effects of improper normalization.

To validate the effectiveness of our method, we apply it to multiple gait recognition models and evaluate its performance under various occlusion scenarios using the OU-MVLP dataset. Experimental results demonstrate that our approach improves recognition accuracy across different occlusion conditions and multiple recognition models

Presentation attacks pose a persistent challenge to secure face recognition. This work explores the use of 3D time-of-flight (ToF) cameras for presentation attack detection (PAD) by combining depth information with remote photoplethysmography (rPPG). A spectral pulse peak prominence ratio (SPR) is introduced as a PAD-feature based on the rPPG signal, alongside motion-based analysis of depth variations. Experiments with 200 subjects and multiple 3D mask attacks demonstrate measurable differences between bona fide presentations and attacks. The study shows that ToF-based rPPG and depth features can support the detection of advanced presentation attacks, while future work will address robustness against motion and generalization across diverse populations.

This work proposes a Presentation of Attack Detection (PAD) on ID card systems, focusing on improving the detection of manual composite attacks. These kinds of attacks are used in a remote verification system to impersonate the identity of one subject in order to obtain benefits. For this purpose, relevant text can be modified, copied, and pasted. Faces and other areas can be manually changed because of the alignment of the originally issued ID card to a fake ID card. This work proposed an “online” methodology to create composite attack images during the training process and shows that using automatic composite attacks may reduce the error rate in the PAD system. To demonstrate it, three experiments were set up that compared the performance of models trained with and without the aforementioned technique on a private dataset and a cross-evaluation dataset.

As biometric recognition systems become increasingly prevalent in security-critical applications, ensuring their robustness against attacks is paramount to maintaining trust and reliability. This is particularly important for systems relying on highly accurate biometric traits such as the iris, for which malicious users may attempt to gain unauthorized access by resorting to artifacts or synthetic samples, i.e., through presentation attacks. This paper explores the capabilities of two state-of-the-art attention-based architectures, namely vision transformers (ViTs) and shifted windows (Swin) transformers, for iris presentation attack detection (PAD). Furthermore, a detailed analysis is performed on a relatively underexplored aspect, specifically the effects of image compression (JPEG and JPEG AI) on data-driven PAD. The obtained results testify to the effectiveness of transformer-based PAD frameworks, whose behaviors can be interpreted by exploring visual maps under degraded image quality. The trade-off between compression efficiency and PAD accuracy is also analyzed.

Face recognition systems are primarily tailored to adult faces, and numerous studies have shown that their performance significantly declines when applied to children. The development of child-specific recognition models is further constrained by the scarcity of publicly available datasets and the ethical and privacy concerns surrounding the collection of children's biometric data. To overcome these limitations, we propose GANDiffKids, a novel framework for synthesizing realistic child face images with diverse intra-class variations. Our approach combines a generative adversarial network (GAN) with a diffusion model -- a strategy that has proven effective for adult face synthesis in prior work. We demonstrate that GANDiffKids can generate a large-scale dataset of synthetic child faces across a range of image qualities. The resulting dataset is benchmarked using a state-of-the-art face recognition model, establishing its potential as a valuable resource for advancing child-specific face recognition research.

With the growing prevalence of biometric recognition technology, the demand for diverse large-scale test datasets is significant, and data collection costs are increasing. Consequently, synthetic images attract attention as a potential solution. However, some synthetic images deviate from target representativeness; even if visually close, subtle alterations can negatively impact biometric recognition accuracy. To address the challenge, this paper introduces a novel Applicability Evaluation framework for synthetic images and biometric engines based on the image quality and score similarity. Synthetic Image Quality Assessment (SIQA) evaluates how closely synthetic images approximate target representativeness. Score Similarity Assessment (SSA) compares biometric engine scores and verifies identity similarity that is crucial for performance testing. Experimental results demonstrate that this framework effectively identified combinations of synthetic images and biometric engines that maintain recognition accuracy while ensuring the reliability of biometric performance testing. It can improve the trustworthiness and efficiency of biometric performance testing.

Multi-face tracking is a challenging problem, as low-quality face images (e.g., partially occluded faces, lateral faces, motion blur, etc.) are prevalent. A multi-face tracker called FaceQSORT mitigates this problem by combining two different types of features (i.e., biometric face features and generic appearance features) extracted from the same image (face) patch. In particular, the appearance features provide a general visual representation of the image (face) patch and compensate for limited biometric information. To extract these appearance features, a generic image classifier is utilized. In this work, the selection of the appearance feature model (image classifier) and the biometric feature model (face recognition model) is evaluated empirically. For this purpose, the following questions are investigated: (i) How crucial is the choice of appearance model? (ii) How does the choice of appearance model affect the inter-person ID switches? To answer these questions, a comprehensive experimental evaluation is conducted that includes multiple tracking parameter settings and different combinations of biometric and appearance model pairs.

Iris segmentation is an important step in determining downstream iris recognition performance. Most iris segmentation systems fail to reliably segment the iris images at extreme off-angles that go beyond 30 degrees. Obtaining manual ground truth for different extreme angles is both a costly and a time-consuming undertaking. In this paper, our aim is to automate this iris segmentation process for different extreme gaze angles, while minimizing the iris recognition performance loss. To address this, we propose an Iris segmentation method using a novel distance-based loss function and a residual-based ellipse fitting method to overcome these obstacles. The novel method shows an overall best performance of 0.9809 Area Under Curve (AUC) and an equal error rate (EER) of 0.0628, reducing the gap to the performance using ideal manual segmentation to 1.88% in terms of AUC and 0.056% in terms of EER.

Academic evaluation of speech technology and AI is currently heavily based on objective quality metrics. In contrast, in the real-world deployment of that technology, such objective metrics are only one small piece in a bigger puzzle. Through a practical use case where we have developed speech technology and AI for a municipal service, this presentation discusses the real-world goals and constraints to the design process, in particular from the perspective of trust and privacy. We find that: Participatory design is a useful tool for scoping challenges and solutions, the often-feared legal challenges are solvable, and trust is a useful concept in discussions but usually a proxy for the real project goals. Such understanding has helped us reframe speech processing challenges to focus on meaningful and rewarding tasks.

As biometric authentication becomes more common, protecting biometric data is becoming increasingly important. One widely used protection method is encryption. However, not all encryption methods are suitable for biometric data. On the one hand, the encryption solution can lead to worse performance or on the other hand, not fulfil all required security measures. This paper proposes a solution for iris template protection that results in hash-encrypted data using Maximum Entropy Binary (MEB) codes and Convolutional Neural Networks (CNN). The method is compared to a baseline approach to demonstrate competitive recognition performance. In addition, the privacy and security properties of the template protection are evaluated.

This paper presents a post-quantum secure multi-party computation (MPC) scheme using homomorphic encryption (PQ-MHE) applied for fingerprint template protection. The proposed scheme leverages the CKKS cryptosystem for homomorphic operations and additive secret sharing (ASS) for splitting, storing, and comparing biometric templates across multiple servers. This architecture ensures that no single entity ever has access to the unencrypted templates. We benchmark the system using two-, three-, and four-server configurations. The scheme achieves a comparison time of 1.24 seconds and incurs 35 MB of communication, which can be reduced by approximately 98% through caching and reusing the CKKS cryptographic context on the server side. The inherent approximate arithmetic of CKKS introduces a minor degradation, resulting in a 0.5 percentage point decrease in the true acceptance rate (TAR) compared to unencrypted fingerprint samples. The system is secure under a semi-honest adversary model with non-colluding servers and complies with the biometric information protection standards outlined in ISO/IEC 24745.

Age verification is increasingly critical for regulatory compliance, user trust, and the protection of minors online. Historically, solutions have struggled with poor accuracy, intrusiveness, and significant security risks. More recently, concerns have shifted toward privacy, surveillance, fairness, and the need for transparent, trustworthy systems. In this paper, we propose Biometric Bound Credentials (BBCreds) as a privacy-preserving approach that cryptographically binds age credentials to an individual's biometric features without storing biometric templates. This ensures only the legitimate, physically present user can access age-restricted services, prevents credential sharing, and addresses both legacy and emerging challenges in age verification. enhances privacy.

This contribution summarizes a pilot study on the feasibility of using a selected commercial electronic fingerprint biometric capture subsystem for post mortem biometric recognition. The study investigates the effectiveness of contact and contactless devices in capturing fingerprints from deceased individuals within 48 hours after death. The environmental conditions, legal and medical implications, and post mortem changes that affect skin integrity are examined. Fingerprint quality and interoperability were evaluated using the NFIQ 2 and IDkit tools. The results show that while traditional dactyloscopic cards offer lower image quality for algorithmic comparison, some modern digital devices provide promising results for post mortem fingerprint acquisition and comparison.

We investigate the capabilities of deep learning methods for the quality assessment of fingerprint samples. Starting from general considerations for the implementation of a deep learning-based fingerprint quality assessment, we modify and fine-tune six convolutional neural networks and one vision transformer for the specific task. For this, we generate a synthetic training database and propose a labelling based on the normalized average sample comparison score.

The obtained results in terms of error vs. discard characteristic curves show that all deep learning methods outperform the NFIQ 2.3 baseline. Most notably, the best-performing VGG16 model results in a superior predictive performance than NFIQ 2.3 on all tested datasets. A detailed interpretation of the obtained results paves the way for more specific and tailored solutions.

We suggest to consider the proposed algorithms as a starting point for a development of a deep learning-based fingerprint quality assessment methods and make them publicly available.

Fingerprints are widely recognized as one of the most unique and reliable characteristics of human identity, making them a preferred choice for biometric-based authentication systems. The use of contactless fingerprints has emerged as an alternative. This paper focuses on the development of a deep learning-based segmentation tool for contactless fingerprint localization and segmentation. Our system leverages deep learning techniques to achieve high segmentation accuracy and reliable extraction of fingerprints from contactless fingerprint images. In our evaluation, our segmentation method demonstrated an average mean absolute error (MAE) of 30 pixels, an error in angle prediction (EAP) of 5.92 degrees, and a labeling accuracy of 97.46%. These results demonstrate the effectiveness of our novel contactless fingerprint segmentation and extraction tools.

Tattoos have been effectively used as soft biometrics to assist law enforcement in identifying perpetrators and victims, as they contain discriminative information and serve as a useful indicator to locate members of a criminal gang or organisation. Due to various privacy issues in the acquisition of images containing tattoos, only a limited number of databases exist. To overcome the handicap, this work presents a robust new tattoo retrieval framework that combines global features and geometrical precision extracted from foundation models for an accurate tattoo representation. The experimental evaluation conducted on a challenging tattoo database reported, in a closed-set protocol, a rank-1 accuracy of 90.91%, outperforming the state-of-the-art TattTRN by a large margin. Similar trends are also achieved for the open-set scenario: the Equal Error Rate is 11.56%, compared to 20.75% yielded by TattTRN. These results demonstrate that combining different foundation models without the need for training can significantly improve accuracy, setting a new benchmark for future tattoo recognition systems.

As biometric authentication becomes increasingly integrated into everyday technology, fingerprint recognition remains a cornerstone due to its reliability and uniqueness. This talk explores the advancements and comparative analysis of contactless 3D fingerprint scanning versus traditional contact-based scanning and smartphone-based 2D contactless scanning. We will discuss the fundamental differences in data acquisition, accuracy, hygiene and user experience among the three modalities. Special focus will be placed on how 3D contactless methods capture depth and fine ridge structure, enabling robust recognition even in challenging scenarios. Through this comparative lens, the talk aims to outline the potential and practical considerations for deploying contactless 3D fingerprint recognition in real-world applications.