German Research Centre for Artificial Intelligence

Abstract Context‐centric proactive information delivery (PID) is a relatively underexplored domain within recommender systems (RS) aimed at enhancing Knowledge Workers' productivity by proactively providing relevant information during digital tasks. These RS anticipate user needs by leveraging personal knowledge modeling, context recognition, and recommendation techniques to deliver timely and relevant resources without requiring explicit searches. Developing such RS faces various challenges in addition to common challenges within the broader RS landscape. These challenges range from data‐related issues, such as handling heterogeneous and noisy data, to user‐centric concerns, including privacy, explainability, and the lack of explicit feedback, as well as system and algorithmic challenges like cross‐application context modeling and scalability. This paper explores the opportunities, challenges, and future directions for PID, outlining key advancements and enabling technologies that support its development.

This paper introduces a pioneering, end-to-end system designed for the accurate estimation of PM2.5 concentrations, which uniquely integrates custom-designed hardware with advanced computational algorithms. A central innovation of this work is the development of a seamless data pipeline that connects real-time vehicle tracking with sophisticated air quality prediction, offering a novel and comprehensive solution for urban environmental monitoring. The system employs a cost-effective, custom-built sensor package, including PMS7003 and BME280 sensors alongside a 5MP camera, with a bespoke hardware design specifically engineered to enhance operational stability. For the estimation component, a rigorous comparative analysis was conducted, demonstrating that the Cubist regression model significantly outperforms other contemporary machine learning and traditional mathematical models; this success is attributed to its superior ability to model the complex, nonlinear relationship between observed traffic density and ambient PM2.5 levels. Furthermore, a streamlined vehicle counting algorithm, leveraging a fine-tuned YOLOv7 model, ensures robust and accurate traffic detection performance across various lighting and environmental conditions. This research successfully establishes an optimal PM2.5 estimation pipeline based on real-world vehicle counts, presenting an integrated framework for dynamic urban air quality analysis. Received: 30 August 2024 | Revised: 30 September 2025 | Accepted: 14 January 2026 Conflicts of Interest The authors declare that they have no conflicts of interest to this work. Data Availability Statement The data that support the findings of this study are openly available in GitHub at https://github.com/comrang-altf4/PM2.5. Author Contribution Statement Chuong Dinh Le: Conceptualization, Methodology, Software, Validation, Formal analysis, Investigation, Data curation, Writing – original draft, Writing – review & editing, Visualization. Hoang Viet Pham: Conceptualization, Methodology, Software, Investigation, Data curation, Writing – original draft, Visualization. Thinh Gia Tran: Conceptualization, Methodology, Software, Validation, Investigation, Writing – original draft, Visualization. An Dinh Le: Conceptualization, Methodology, Resources, Writing – review & editing, Supervision, Project administration. Anh-Duy Pham: Conceptualization, Methodology, Resources, Supervision, Project administration. Dat Thanh Vo: Software, Writing – original draft, Visualization. Hien Bich Vo: Resources, Supervision. Dzung Huy Han: Supervision.

Forest disturbances are among the main drivers of global carbon emissions. These disturbances are associated with various human-related and natural drivers, including unsustainable resource extraction, fires, overgrazing and extreme weather events. Such disturbances vary in intensity and type, ranging from stand-replacing disturbances following clear-cuts or windthrow to more scattered disturbance patterns involving standing deadwood resulting from drought-induced mortality or small-scale canopy removal from selective logging. In recent years, multiple Earth observation products have been generated from Landsat and Sentinel missions to monitor such disturbances. These products vary in their methodological approaches and in their global and temporal coverage. However, there are currently no consistent benchmarks with which to evaluate their performance under different disturbance regimes and drivers. This study aims to evaluate and compare the accuracy and operational applicability of satellite-based forest disturbance products. We compared eight large-scale satellite products for detecting various forest disturbances, such as scattered tree mortality, large-scale removal and natural hazards. The disturbance products compared include Global Forest Change (GFC), DIST-ALERT, DeadTrees.Earth and the European Forest Disturbance Atlas (EFDA), amongst others. The products were compared qualitatively and quantitatively using reference data on disturbance events obtained from globally distributed aerial imagery acquired using unmanned aircraft systems (UAS). We use a total of 35 aerial orthomosaics acquired between 2015 and 2024, obtained from the DeadTrees.Earth platform. We identify forest disturbance types and quantify their extent using visual interpretation. This study advances our understanding of the strengths and limitations of current forest disturbance products by systematically assessing their performance across diverse disturbance types and environmental contexts.

Routine clinical imaging of the retina using optical coherence tomography (OCT) is performed with large slice spacing, resulting in highly anisotropic images and a sparsely scanned retina. Most learning-based methods circumvent the problems arising from the anisotropy by using 2D approaches rather than performing volumetric analyses. These approaches inherently bear the risk of generating inconsistent results for neighboring B-scans. For example, 2D retinal layer segmentations can have irregular surfaces in 3D. Furthermore, the typically used convolutional neural networks are bound to the resolution of the training data, which prevents their usage for images acquired with a different imaging protocol. Implicit neural representations (INRs) have recently emerged as a tool to store voxelized data as a continuous representation. Using coordinates as input, INRs are resolution-agnostic, which allows them to be applied to anisotropic data. In this paper, we propose two frameworks that make use of this characteristic of INRs for dense 3D analyses of retinal OCT volumes. 1) We perform inter-B-scan interpolation by incorporating additional information from en-face modalities, that help retain relevant structures between B-scans. 2) We create a resolution-agnostic retinal atlas that enables general analysis without strict requirements for the data. Both methods leverage generalizable INRs, improving retinal shape representation through population-based training and allowing predictions for unseen cases. Our resolution-independent frameworks facilitate the analysis of OCT images with large B-scan distances, opening up possibilities for the volumetric evaluation of retinal structures and pathologies. Our code is available at <a href='https://github.com/tkepp/ResA-OCT'>https://github.com/tkepp/ResA-OCT</a>

Abstract Recent advances in vision-language models (VLMs) have significantly improved visual emotion recognition using multimodal contextual reasoning. However, their deployment remains limited by heavy computational demands and centralized data requirements, limiting accessibility for privacy-sensitive or resource-limited settings. This work introduces CAREFL, a context-aware and federated framework that combines the reasoning capabilities of large foundation VLMs with the efficiency of compact, quantized models. In CAREFL, rich textual context is first generated from a powerful VLM and then used to guide the local fine-tuning of a small model under a federated learning setup. This approach enables distributed adaptation of multimodal emotion understanding while minimizing communication cost and preserving data privacy. Experiments on EMOTIC and CAER-S datasets demonstrate that CAREFL achieves up to 40.20\% recall and 55.40\% F1-score, surpassing heavier centralized baselines while operating within a 6 GB memory budget. The results suggest that leveraging contextual reasoning from foundation models can unlock scalable, privacy-preserving emotion recognition across heterogeneous edge environments.

Biometric authentication has long been regarded as a foundational element of identity verification, leveraging unique physiological and behavioral traits to enhance security beyond traditional passwords. While it offers notable advantages such as convenience and resistance to identity theft, concerns are mounting regarding privacy, susceptibility to spoofing, and the irreversibility of compromised biometric identifiers. These weaknesses are becoming increasingly critical as digital infrastructures evolve into distributed, dynamic environments in which static trust models are no longer sufficient. Moreover, several traditional modalities- such as fingerprints, iris scans, and voice recognition- have already been breached. However, Artificial Intelligence (AI) methods are reshaping this landscape by introducing adaptive and context‑aware features into biometric systems. Machine Learning (ML) techniques enhance accuracy, enable continuous authentication, and support multimodal fusion, while anomaly‑detection mechanisms improve resilience against sophisticated attacks. Generative AI (GenAI) plays a particularly significant role, though it introduces a paradox: it empowers defenders through realistic attack simulations and robustness testing, yet simultaneously equips attackers with tools for producing deepfakes and synthetic identities, thereby expanding the attack surface. In this evolving security landscape, Zero‑Trust Architectures (ZTA) have gained prominence as a model that replaces assumptions of inherent trust with continuous verification mechanisms. The use of biometric data within ZTA can enhance the reliability of identity verification; however, it also intensifies several existing issues. Biometric identifiers must be handled and stored in ways that safeguard individual privacy and align with relevant legal requirements, and the incorporation of AI‑based assessment methods introduces additional concerns regarding potential bias, transparency, and oversight. Moreover, combining AI‑supported biometric systems with Zero‑Trust principles raises further questions about scalability, system compatibility, and the broader ethical consequences of more pervasive identity monitoring. This work therefore examines the convergence of biometrics, AI, and Zero‑Trust principles from a critical perspective. It highlights the dual role of AI as both a source of innovation and a generator of new threats, while identifying opportunities for adaptive security, real‑time threat detection, and improved user experience. By analyzing technical and operational dimensions, the work proposes a roadmap for integrating biometrics into ZTA that balances innovation with accountability and supports trustworthy, resilient cybersecurity frameworks.

Trust is one of the fundamental necessities of human beings and, according to Stephen R. Covey, not simply the “glue of life”, but also the “most essential ingredient in effective communication”. Even though the principles and importance of trust are as old as humanity itself -in ancient times, trusting strangers could mean the difference between life and death, and thus pose an immediate threat to one's social tribe- trust is gaining increasing attention, particularly in light of tomorrow's all-electric society and the decentralization and globalization associated with it. Contrary to previous decades, physical proximity is no longer necessary to access a system; access is possible (almost) anytime, (almost) anywhere. However, trust is a multidimensional concept depending on a multitude of aspects, such as the specific application, the value of the resource, and the available technology, but also -for instance- on the people who are managing access to systems and applications. As societies become increasingly aware of data breaches, algorithmic surveillance, and commercialization of personal data, cultural values have shifted toward individual agency and distrust of centralized institutions. As trust could never be taken for granted it rather must be earned, orchestrated and continuously verified. Technologies like multi-factor authentication (MFA) reflect this, offering multi-layered mechanisms for identity verification and access protection. These developments illustrate a broader societal reshaping of trust -from implicit trust toward conditional, data-driven security- and show how technological design is shaped by evolving social anxieties and expectations. Against this background, this work focuses on socio-cultural influences on the definition of trust and trustworthiness, such as origin (geography, culture, political system) or educational background and training. In particular, it follows research questions including: i) How do socio-cultural perceptions of trust and identity shape the development and adoption of digital security technologies such as MFA; ii) in what/which ways does the rise of digital surveillance and data breaches influence societal expectations of privacy and trust in technological systems?; and iii) how is the concept of trust redefined in the digital age, and what role do authentication technologies play in mediating this transformation?