Daniel DeAlcala

PhD Candidate in Artificial Intelligence

Email Scholar GitHub LinkedIn CV
Photo of Daniel DeAlcala

About

I’m a PhD candidate in Artificial Intelligence at the Universidad Autónoma de Madrid, advised by Julian Fierrez and Aythami Morales, and currently a visiting researcher with the Docling team at IBM Research Zurich.

I work on vision-language models for documents: turning pages into structured text and code, and telling which parts of a document were written or edited by AI. My PhD is about auditing AI models: I introduced the Membership Inference Test (MINT), which checks whether given data was used to train a model.

So far: 16 peer-reviewed papers, 9 as first author, including ICCV and IEEE Access, and a patent application filed by IBM. When I’m not training models, I’m probably traveling, watching football, or climbing.

  • Vision-Language Models
  • Document Intelligence
  • Image-to-Code Generation
  • AI-Generated Content Detection
  • Membership Inference
  • Trustworthy AI
  • Computer Vision

News

  • Starting a research stay at IBM Research – Zurich with the Docling team (Sep–Dec 2026). :mountain:
  • A busy 2026 so far :tada: Papers at IEEE CAI 2026 (gradient-based MINT for LLMs), IEEE COMPSAC 2026 (MINT Demo 2, now for vision-language data) and two at the ICDAR 2026 Workshops, plus the LoRA-MINT preprint.
  • Completed my Research Internship at IBM Research
  • Just attended ICCV, an unforgettable experience where we presented 3 papers and I met incredibly interesting people, as always happens at these conferences :smile:
  • Paper “Is My Data in Your AI? Membership Inference Test (MINT) applied to Face Biometrics” accepted on IEEE Access!
All news

Featured projects

ReTeX

In progress

A vision-language model that reads the image of a page and writes LaTeX that compiles back into the same page. 0.74 mean fidelity on 400 unseen arXiv pages.

VLMImage-to-codeLaTeX

DocAI

Under review · ICLR 2027

Detecting and localizing AI-edited content in whole documents, with a 35K-page benchmark and a 256M-parameter VLM detector. Built at IBM Research.

Document forensicsVLM

AI-MINTest

Live

A public web tool that checks whether given images or texts were used to train an AI model, based on the Membership Inference Test from my PhD.

AuditingPrivacyDemo

Selected publications

  1. Under review
    DocAI: Detecting and Localizing AI-Edited Content in Documents
    Daniel DeAlcala, Ching-Yun Ko, Aythami Morales, Julian Fierrez, Ruben Tolosana, and Pin-Yu Chen
    Under review at the International Conference on Learning Representations (ICLR 2027), 2026

    TL;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.

    Paper Code Cite
  2. ICCV 2025
    Active Membership Inference Test (aMINT): Enhancing Model Auditability with Multi-Task Learning
    Daniel DeAlcala, Aythami Morales, Julian Fierrez, Gonzalo Mancera, Ruben Tolosana, and Javier Ortega-Garcia
    Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2025

    TL;DRActive MINT trains the model and its auditor together, so training-data membership can be verified later: above 80% accuracy from MobileNet to ViT.

    Paper Code Cite
  3. IEEE Access
    Is My Data in Your AI? Membership Inference Test (MINT) applied to Face Biometrics
    Daniel DeAlcala, Aythami Morales, Julian Fierrez, Gonzalo Mancera, Ruben Tolosana, and Javier Ortega-Garcia
    IEEE Access, 2025

    TL;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.

    Paper Code Cite
  4. COMPSAC 2023
    Measuring Bias in AI Models: An Statistical Approach Introducing N-Sigma
    Daniel DeAlcala, Ignacio Serna, Aythami Morales, Julian Fierrez, and Javier Ortega-Garcia
    2023 IEEE 47th Annual Computers, Software, and Applications Conference (COMPSAC)

    TL;DRN-Sigma, a statistical method from physics, applied to measuring bias in face recognition models.

    Paper Cite
All publications

Experience

  1. IBM Research – Zurich

    Sep – Dec 2026
    Visiting Researcher, Docling team · hosted by Peter Staar Zurich, Switzerland

    Contributing to Docling, IBM's open-source document conversion toolkit (60K+ GitHub stars), and developing a VLM that links the fields of a form to their values.

  2. IBM Research · MIT-IBM Watson AI Lab

    Sep – Dec 2025
    Visiting Researcher · hosted by Pin-Yu Chen New York, USA

    Built DocAIgen, a 35K-page benchmark of AI-edited documents, and DocAIdet, a 256M-parameter VLM that detects and segments the edited regions, above GPT-5.2, Gemini 3 Flash and Claude Opus 4.5 used as detectors. IBM filed a patent application on the method.

  3. BiometricsAI Lab, Universidad Autónoma de Madrid

    Jan 2022 – present
    PhD Researcher and FPU Fellow · advised by Julian Fierrez and Aythami Morales Madrid, Spain

    Introduced the Membership Inference Test (MINT) for auditing the training data of image and text models.

Education

  1. Ph.D. in Telecommunications and Computer Engineering

    2022 – expected early 2027
    Universidad Autónoma de Madrid

    Thesis: Auditing Artificial Intelligence: Membership Inference and Methods for Transparent, Fair and Trustworthy AI.

  2. M.Sc. in Deep Learning for Audio and Video Signal Processing

    2020 – 2021
    Universidad Autónoma de Madrid

  3. B.Sc. in Telecommunications Engineering

    2016 – 2020
    Universidad Autónoma de Madrid

    Ranked 1st in the graduating class, with 8 Matrículas de Honor.