publications
My published research. Hundreds of hours of work distilled into a few pages of "simple" PDFs.
2026
-
Under reviewDocAI: Detecting and Localizing AI-Edited Content in DocumentsUnder review at the International Conference on Learning Representations (ICLR 2027), 2026TL;DRDocAIgen rewrites paragraphs, figures and tables of real PDFs with any generative model to build a 35K-page benchmark; DocAIdet, a 256M-parameter VLM, detects and localizes the AI-edited regions.
-
IEEE CAI 2026Is My Text in Your AI Model? Gradient-based Membership Inference Test applied to LLMsProceedings of the IEEE Conference on Artificial Intelligence (CAI), 2026TL;DRGradient-based membership inference for language models: across 7 Transformer models and 2.5M sentences, AUC 0.85 to 0.99.
-
COMPSAC 2026Is My Vision-Language Data in Your AI? Membership Inference Test (MINT) Demo 2IEEE Annual Computers, Software, and Applications Conference (COMPSAC), 2026TL;DRA public web platform that audits image and text models (MINT, aMINT, gMINT), with up to 90% accuracy in detecting training data.
- ICDAR-W 2026Comparative Study of Domain-adapted VLMs for General Document Visual Question AnsweringInternational Conference on Document Analysis and Recognition (ICDAR) Workshops, 2026
TL;DRA comparison of domain-adapted vision-language models for general document visual question answering.
- ICDAR-W 2026CrimeNER Demo: Named-Entity Recognition in the Crime DomainInternational Conference on Document Analysis and Recognition (ICDAR) Workshops, 2026
TL;DRA demonstrator for named-entity recognition in crime-domain documents.
2025
-
ICCV 2025Active Membership Inference Test (aMINT): Enhancing Model Auditability with Multi-Task LearningProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2025TL;DRActive MINT trains the model and its auditor together, so training-data membership can be verified later: above 80% accuracy from MobileNet to ViT.
-
ICCV 2025DeepID Challenge of Detecting Synthetic Manipulations in ID DocumentsProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2025TL;DRChallenge on detecting synthetic manipulations in ID documents. Our UAM team had the best score of 26 teams on the organisers’ private test set of 20K real ID documents.
-
IEEE AccessIs My Data in Your AI? Membership Inference Test (MINT) applied to Face BiometricsIEEE Access, 2025TL;DRMembership Inference Test (MINT) for face recognition: using model activations, it detects whether a face was in the training set with up to 90% accuracy on 22M+ images.
-
CVPR-W 2025gMINT: Gradient-based Membership Inference Test applied to Image ModelsProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2025TL;DRgMINT audits image models from per-sample gradient updates and reaches near-perfect training-data detection in most settings.
- AAAI-W 2025Membership Inference Test: Auditing Training Data in Object Classification ModelsAAAI Workshop on Deployable AI, 2025
TL;DRMembership Inference Test applied to auditing the training data of object classification models.
2024
-
CVPR-W 2024A Comprehensive Analysis of Factors Impacting Membership InferenceProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2024TL;DRWhich training factors make membership easier or harder to detect: number of epochs, loss function and dropout, studied on 22M+ face images.
2023
-
CVPR-W 2023BeCAPTCHA-Type: Biometric Keystroke Data Generation for Improved Bot DetectionProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2023TL;DRSynthetic keystroke data generation to train better bot detectors, on 136M keystroke events from 168K subjects.
2022
-
AAAI-W 2022ALEBk: Feasibility Study of Attention Level Estimation via Blink Detection applied to e-LearningAAAI Workshop on Artificial Intelligence for Education (AI4EDU), 2022TL;DREstimating the attention level of students in online sessions from eye blink frequency, with a CNN blink detector.
