Our Work

The number of IoT devices continues to rise and it is imperative to detect anomalies in real-time to ensure the security, reliability and efficiency of the system. Current deep neural network models h…

The number of IoT devices continues to rise and it is imperative to detect anomalies in real-time to ensure the security, reliability and efficiency of the system. Current deep neural network models have great memory and computing requirements, so it is not easy to use on the edge device. The benefit of federated learning is that it leaves raw data on the individual device, thus improving privacy. Nonetheless, efficient feature learning by compressing the size of quantum datasets has not been built into current federated learning frameworks, and the efficacy of quantum-based approaches under realistic noise in the hardware have not been widely investigated. To address privacy concerns, QEdgeNet is a lightweight hybrid quantum-classical federated learning (QLFL) model for anomaly detection in edge-IoT networks . In this model, the features of network traffic are first compressed by principal component analysis, and then the network traffic is mapped to a compact 4-qubit variational quantum circuit for the classification of anomalous traffic. This model is an aggregated model updated using the Cheon-Kim-Kim-Song (CKKS) homomorphic encryption mechanism, but without exposing the raw network data of clients, and raw traffic data never escapes from a client device during the training process. The applicability of the suggested approach in an idealized simulator is quantified by measuring the inference performance in a simulator simulated quantum hardware noise, the accuracy loss from the hardware noise is recovered by extrapolating to zero noise.

Introduction The increasing complexity of cyber threats in Internet of Things (IoT) environments requires the development of efficient and privacy preserving Intrusion Detection Systems (IDS). IoT net…

Introduction The increasing complexity of cyber threats in Internet of Things (IoT) environments requires the development of efficient and privacy preserving Intrusion Detection Systems (IDS). IoT networks generate high-dimensional and heterogeneous traffic data, making effective feature selection and temporal attack detection challenging, particularly in distributed and resource-constrained environments.Objectives: This research has the following objectives: (1) to develop a hybrid feature selection method using Hybrid Grey Wolf Optimization integrated with Cuckoo Search and chaotic Lévy flight (HGS-CS), (2) to reduce the dimensionality of IoT network traffic while retaining highly discriminative features, (3) to develop a TSTformer-LSTM model combining a Transformer Encoder, Bidirectional LSTM, and Gated Fusion mechanism for capturing global contextual and temporal dependencies, and (4) to develop a privacy-preserving federated IDS using SwarmFed with attention-based aggregation, Stable Focal Loss, and centralized fine-tuning.Methods: This research proposed a novel framework of hybrid feature selection using Hybrid Grey Wolf Optimization integrated with Cuckoo Search enhanced with chaotic Levy flight (HGS-CS) and a TSTformer-LSTM model integrated with a Transformer Encoder with a Bidirectional LSTM and a Gated Fusion mechanism that captures both the global dependencies in context and also the temporal dependencies which validated on the datasets UNSW-NB15 and BoTNeTIoT-L01. The HGS-CS performs the multi-objective feature selection which results in a dimensionality reduction of 86.96% by selecting only 3 highly discriminative features in BoTNeTIoT-L01 dataset and 8 features in UNSWNB-15 dataset. The model is trained based on a SwarmFed federated learning framework with attention-based aggregation, Stable Focal Loss and centralized fine-tuning.Results: Experimental results show performance which achieves an accuracy of 96.9% on BoTNeTIoT-L01 dataset and 89% on UNSWNB-15 dataset. The proposed framework provides a robust, efficient, and privacy-aware IDS solution suitable for distributed and resource-constrained IoT environments

Título: Los Modelos Mentales: El Software que Ejecuta tu Empresa sin que te Des Cuenta   Autor: Carlos Guillermo Gallardo Heredia   Año: 2026   RESUMEN   Las decision…

Título: Los Modelos Mentales: El Software que Ejecuta tu Empresa sin que te Des Cuenta   Autor: Carlos Guillermo Gallardo Heredia   Año: 2026   RESUMEN   Las decisiones que parecen "lógicas" para el departamento de ventas resultan "absurdas" para el de operaciones. Los conflictos entre áreas no surgen de mala voluntad, sino de modelos mentales distintos: suposiciones profundamente arraigadas que determinan cómo cada persona entiende el mundo y cómo actúa. Este artículo explora el concepto de modelos mentales desde la perspectiva de Peter Senge y del pensamiento sistémico aplicado, y propone una herramienta concreta —el salto de abstracción— para hacer explícitos esos modelos y alinear a los equipos sin necesidad de que todos piensen igual. Se integran los conceptos del libro SISP™, el libro De Caos a Autonomía y el manuscrito AMSI (metodología participante en el Premio Chihuahua 2026). Se incluye una comparación explícita entre el consultor lineal y el consultor de la complejidad frente al trabajo con modelos mentales.   Palabras clave: modelos mentales, Peter Senge, salto de abstracción, conflicto organizacional, sistemas complejos, PYMES, AMSI, SISP™, consultoría sistémica.   ABSTRACT   Decisions that seem "logical" to the sales department are "absurd" to the operations department. Conflicts between areas do not arise from ill will, but from different mental models: deeply rooted assumptions that determine how each person understands the world and how they act. This article explores the concept of mental models from the perspective of Peter Senge and applied systems thinking, and proposes a concrete tool —the ladder of inference— to make those models explicit and align teams without requiring everyone to think alike. Concepts from the book SISP™, the book From Chaos to Autonomy, and the AMSI manuscript (methodology participating in the Premio Chihuahua 2026) are integrated. An explicit comparison between the linear consultant and the complexity consultant in working with mental models is included.   Keywords: mental models, Peter Senge, ladder of inference, organizational conflict, complex systems, SMEs, AMSI, SISP™, systemic consulting.   INFORMACIÓN ADICIONAL   Este trabajo se fundamenta en la arquitectura de decisión en sistemas complejos, integrando los marcos AMSI (Arquitectura Metodológica para Análisis e Intervención en Sistemas Organizacionales Complejos) y SISP™ (Sistema Integral para la comprensión, decisión e intervención en sistemas complejos), desarrollados por el autor. Forma parte de una serie de publicaciones sobre pensamiento sistémico aplicado a la gestión empresarial y la consultoría organizacional.   LICENCIA   Creative Commons Atribución-NoComercial-CompartirIgual 4.0 Internacional (CC BY-NC-SA 4.0).   © 2026 Carlos Guillermo Gallardo Heredia

Code and processed neighborhood-level data to reproduce the tables and figures of the article 'The Wage Dividend of Pacification: Violence Trajectories and Neighborhood Wages in Medellín, 2006–2018'. …

Code and processed neighborhood-level data to reproduce the tables and figures of the article 'The Wage Dividend of Pacification: Violence Trajectories and Neighborhood Wages in Medellín, 2006–2018'. ECV microdata are not redistributed and are available from the Alcaldía de Medellín.

Ride-hailing apps are built mainly to book a ride, match a driver and reach the destination quickly. Safety and accessibility are usually added afterwards as a separate SOS button or a support-ticket …

Ride-hailing apps are built mainly to book a ride, match a driver and reach the destination quickly. Safety and accessibility are usually added afterwards as a separate SOS button or a support-ticket system, so women, elderly riders and persons with disabilities (PWD) are often poorly served. This paper presents SafeGo, a ride-hailing platform that brings passenger-specific ride modes, machine-learning-based safety-risk prediction, accessibility-aware driver matching, PIN-based ride verification, live monitoring, SOS emergency response, transparent fare calculation and an administrative command center into one system. SafeGo provides four ride modes (Normal, Pink, Elderly and PWD), and a Random Forest classifier uses temporal, geographical, trip-level and mode-related features to place each trip in the Stable, Cautious or High Priority category. On a held-out synthetic test set of 2,000 samples the classifier achieved 99.80% accuracy, 99.87% macro precision, 99.78% macro recall and 99.82% macro F1-score, with a mean prediction time of 32.27 ms. Scenario-based tests show that predictions change with the ride mode and trip context. Because the dataset is synthetic, these figures describe model behaviour on the evaluation data and are not a claim about real-world accuracy

The Waveform Norm reduction method was introduced into SAE ARP5415 Revision B. This method allows for all aircraft response waveforms to be reduced to four waveform norms that can be used to select st…

The Waveform Norm reduction method was introduced into SAE ARP5415 Revision B. This method allows for all aircraft response waveforms to be reduced to four waveform norms that can be used to select standard test waveforms. Use of this method can greatly reduce the processing time for waveform reduction via automated processing methods. Revision B was issued in 2020. This paper provides feedback on the waveform norms method, and discusses some of the authors’ experiences with nuances surrounding peak rate of change calculations. The paper will also present alternative methods for calculating peak rate of change and provide a procedure for relating peak rate of change to specific standard test waveforms.

The management problem addressed in this study is that fragmented digital-marketing analytics cannot jointly process heterogeneous high-volume data, adapt predictions to changing customer behavior, ex…

The management problem addressed in this study is that fragmented digital-marketing analytics cannot jointly process heterogeneous high-volume data, adapt predictions to changing customer behavior, explain model recommendations, and optimize financial and engagement outcomes. This limitation reduces managerial trust and makes ROI-oriented decisions difficult in omnichannel environments. The study proposes a Hybrid Explainable Artificial Intelligence and Big Data Architecture (HEAIBDA) that integrates distributed data processing, Random Forest, XGBoost, Long Short-Term Memory models, SHAP/LIME explanations, and a composite Digital Marketing Performance Index (DMPI). An explanatory, longitudinal, comparative multiple-case design is used with harmonized secondary indicators for ten technology-oriented corporations during 2020–2025. The evaluation compares a 2020–2022 baseline with a 2023–2025 HEAIBDA-aligned optimization scenario. Mean ROI increases from 17.20% to 26.13%, conversion rate from 3.31% to 4.88%, customer engagement from 0.659 to 0.832, and DMPI from 0.628 to 0.818, while customer acquisition cost decreases from USD 37.00 to USD 25.10. Prediction transparency rises from 47.66% to 83.61%, and all ten cases move in the expected direction. The study concludes that combining scalable analytics, hybrid prediction, and explainability offers a coherent decision-support architecture for multidimensional marketing performance management. Because the evidence is based on secondary aggregate data and scenario evaluation rather than controlled field deployment, the findings demonstrate design plausibility and cross-case consistency, not causal impact.

Precision agriculture requires image-analysis models that remain reliable when crop type, phenological stage, environment, and sensing platform change, yet recent supervised and self-supervised studie…

Precision agriculture requires image-analysis models that remain reliable when crop type, phenological stage, environment, and sensing platform change, yet recent supervised and self-supervised studies commonly address a single image source or a single downstream task. This study presents AgriDA-SSL, a domain-adaptive self-supervised framework that learns from unlabeled satellite, UAV, and field-level imagery and is fine-tuned with limited annotations. The framework combines a hybrid CNN-transformer encoder with contrastive representation learning, masked image reconstruction, prototype-guided clustering, domain-consistency regularization, and an adaptive augmentation policy. The evaluation covers crop classification, disease recognition, field segmentation, and yield-related visual estimation across PlantVillage, BigEarthNet/Sentinel-2, Agriculture-Vision, LUCAS crop images, and UAV imagery. Under the reported experimental protocol, AgriDA-SSL attains 98.91% accuracy, 98.64% precision, 98.52% recall, 98.58% F1-score, and 0.993 ROC-AUC for the principal classification experiment, while ablation and low-label analyses indicate that each major component contributes to performance. The study's contribution is therefore not a claim that self-supervision alone is new, but the integration and evaluation of complementary self-supervised objectives, cross-domain regularization, and adaptive augmentation within one multi-source, multi-task agricultural vision pipeline. The results support label-efficient agricultural monitoring, while the reported threats to validity delimit the extent to which benchmark performance can be generalized to unseen farms, sensors, seasons, and deployment conditions.Precision agriculture requires image-analysis models that remain reliable when crop type, phenological stage, environment, and sensing platform change, yet recent supervised and self-supervised studies commonly address a single image source or a single downstream task. This study presents AgriDA-SSL, a domain-adaptive self-supervised framework that learns from unlabeled satellite, UAV, and field-level imagery and is fine-tuned with limited annotations. The framework combines a hybrid CNN-transformer encoder with contrastive representation learning, masked image reconstruction, prototype-guided clustering, domain-consistency regularization, and an adaptive augmentation policy. The evaluation covers crop classification, disease recognition, field segmentation, and yield-related visual estimation across PlantVillage, BigEarthNet/Sentinel-2, Agriculture-Vision, LUCAS crop images, and UAV imagery. Under the reported experimental protocol, AgriDA-SSL attains 98.91% accuracy, 98.64% precision, 98.52% recall, 98.58% F1-score, and 0.993 ROC-AUC for the principal classification experiment, while ablation and low-label analyses indicate that each major component contributes to performance. The study's contribution is therefore not a claim that self-supervision alone is new, but the integration and evaluation of complementary self-supervised objectives, cross-domain regularization, and adaptive augmentation within one multi-source, multi-task agricultural vision pipeline. The results support label-efficient agricultural monitoring, while the reported threats to validity delimit the extent to which benchmark performance can be generalized to unseen farms, sensors, seasons, and deployment conditions.Precision agriculture requires image-analysis models that remain reliable when crop type, phenological stage, environment, and sensing platform change, yet recent supervised and self-supervised studies commonly address a single image source or a single downstream task. This study presents AgriDA-SSL, a domain-adaptive self-supervised framework that learns from unlabeled satellite, UAV, and field-level imagery and is fine-tuned with limited annotations. The framework combines a hybrid CNN-transformer encoder with contrastive representation learning, masked image reconstruction, prototype-guided clustering, domain-consistency regularization, and an adaptive augmentation policy. The evaluation covers crop classification, disease recognition, field segmentation, and yield-related visual estimation across PlantVillage, BigEarthNet/Sentinel-2, Agriculture-Vision, LUCAS crop images, and UAV imagery. Under the reported experimental protocol, AgriDA-SSL attains 98.91% accuracy, 98.64% precision, 98.52% recall, 98.58% F1-score, and 0.993 ROC-AUC for the principal classification experiment, while ablation and low-label analyses indicate that each major component contributes to performance. The study's contribution is therefore not a claim that self-supervision alone is new, but the integration and evaluation of complementary self-supervised objectives, cross-domain regularization, and adaptive augmentation within one multi-source, multi-task agricultural vision pipeline. The results support label-efficient agricultural monitoring, while the reported threats to validity delimit the extent to which benchmark performance can be generalized to unseen farms, sensors, seasons, and deployment conditions.Precision agriculture requires image-analysis models that remain reliable when crop type, phenological stage, environment, and sensing platform change, yet recent supervised and self-supervised studies commonly address a single image source or a single downstream task. This study presents AgriDA-SSL, a domain-adaptive self-supervised framework that learns from unlabeled satellite, UAV, and field-level imagery and is fine-tuned with limited annotations. The framework combines a hybrid CNN-transformer encoder with contrastive representation learning, masked image reconstruction, prototype-guided clustering, domain-consistency regularization, and an adaptive augmentation policy. The evaluation covers crop classification, disease recognition, field segmentation, and yield-related visual estimation across PlantVillage, BigEarthNet/Sentinel-2, Agriculture-Vision, LUCAS crop images, and UAV imagery. Under the reported experimental protocol, AgriDA-SSL attains 98.91% accuracy, 98.64% precision, 98.52% recall, 98.58% F1-score, and 0.993 ROC-AUC for the principal classification experiment, while ablation and low-label analyses indicate that each major component contributes to performance. The study's contribution is therefore not a claim that self-supervision alone is new, but the integration and evaluation of complementary self-supervised objectives, cross-domain regularization, and adaptive augmentation within one multi-source, multi-task agricultural vision pipeline. The results support label-efficient agricultural monitoring, while the reported threats to validity delimit the extent to which benchmark performance can be generalized to unseen farms, sensors, seasons, and deployment conditions.

Track Changes: studying digital writing processes This keystroke logging application was developed within the project Track Changes. Track Changes: Textual scholarship and the challenge of digital lit…

Track Changes: studying digital writing processes This keystroke logging application was developed within the project Track Changes. Track Changes: Textual scholarship and the challenge of digital literary writing (2018 - Summer 2024) applied new methods, including keystroke logging, to gain insight into the literary writing process of today, which largely takes place on the computer. The aim was to learn what traces the digital work process leaves behind, and how we can use modern techniques to document and analyse the writing process. Twelve Dutch and Flemish authors participated by documenting their writing processes using Inputlog. Keywords: Keystroke logging; writing studies; translation studies; LibreOffice; cognitive processes Project team PI of the project is Karina van Dalen-Oskam (Huygens, KNAW). Dirk van Hulle, Vincent Neijt (both from the Centre for Manuscript Genetics), Mariëlle Leijten and Luuk van Waes (both from Writing and Professional Communication) form the Antwerp part of the team. Within the project, Lamyk Bekius and Floor Buschenhenke were PhD-students. Lamyk Bekius defended her PhD-thesis Behind the computer screens: The use of keystroke logging for genetic criticism applied to born-digital works of literature on 6th October 2023. Floor Buschenhenke's thesis, Entering stories: decoding born-digital fiction writing through keystroke logging, was published in January 2025.  Track Changes is financed by the Humanities Free Competition of the NWO. The project is a collaboration between Huygens Instituut and the University of Antwerp. Keystroke logging Inputlog can be used to study digital writing. The pre-existing application contains both a keystroke registration tool as well as pre- and post-processing options and a range of analyses. To facilitate easier and more accessible logging, the current variant of Inputlog was developed. It is a Libre Office extension. Its sole purpose is to record the writing process, but its output is readable by its big brother, the stand-alone Inputlog application. The extension itself is downloadable from its LibreOffice extensions library page. Developers Inputlog Libre was developed by Resoftlabs: Faruk Diblen, Jisk Attema and Jason Maassen. Since 2026, adaptations are made by internstudents of UAntwerpen.