↓ Skip to main content

Dr. Gabriel Loiseau: PhD Defended!

·394 words·2 mins

On November 2, 2026, Gabriel Loiseau defended his PhD thesis, Privacy and Utility-Preserving Transfer Learning for Text Anonymization. Congratulations, Dr. Loiseau! 🎓

A PhD is a long road, and an industrial one (a CIFRE) adds a second set of expectations: the work has to hold up in front of a scientific jury and be useful to a company. Gabriel managed both, and I’m very proud to have supervised him alongside Marc Tommasi, Damien Sileo and Maxime Meyer.

In Gabriel’s words: “It was an honor to present this work to an outstanding jury. I’m deeply grateful to Pierre Lison, Aurélie Névéol, Ivan Habernal and Claire Gardent for their careful reading and for such a rich discussion.”

The thesis in a nutshell
#

Anonymizing text is not only about removing names. Style, context and rare details are often enough to tell who wrote a text or who it talks about, and aggressive rewriting destroys what made the text useful in the first place. Gabriel’s thesis tackles that trade-off: how to rewrite text so that it protects people while staying useful for the task it was collected for, and how to measure both sides honestly.

Publications
#

Right between his internship and the start of his PhD, Gabriel published his first paper:

  • WikiPhish: A Diverse Wikipedia-Based Dataset for Phishing Website Detection. ACM CODASPY 2024. [link]

Then, during his PhD:

  • TAROT: Task-Oriented Authorship Obfuscation Using Policy Optimization Methods. PrivateNLP 2025 (NAACL workshop). [link]
  • Tau-Eval: A Unified Evaluation Framework for Useful and Private Text Anonymization. EMNLP 2025, Demo Track. [link]
  • Adaptive Text Anonymization: Learning Privacy-Utility Trade-offs via Prompt Optimization. Findings of ACL 2026. [link]
  • Distilling Human-Aligned Privacy Sensitivity Assessment from Large Language Models. LEGAL / CALD-pseudo 2026 (LREC workshop). [link]

As if his days were not busy enough, he also co-authored a paper with fellow PhD student Bastien Liétard, and contributed to Humanity’s Last Exam, a benchmark of expert-level questions for AI models:

  • CALE: Concept-Aligned Embeddings for Both Within-Lemma and Inter-Lemma Sense Differentiation. Bastien Liétard, Gabriel Loiseau. EACL 2026. [link]
  • A benchmark of expert-level academic questions to assess AI capabilities (Humanity’s Last Exam). Nature, 2026. [link]

What’s next
#

Gabriel stays with us at Hornetsecurity, as a Research Engineer in the PAIR (Prototyping, AI and Research) team. I’m glad to keep working with him, and I can’t wait to see the research he’ll bring into production.

Congratulations again, Gabriel!

Damien Riquet
Author
Damien Riquet
Turning NLP research into email protection that runs at scale, and supervising the PhDs behind it.
Lead Research Engineer at Hornetsecurity, a Proofpoint company · AI SIG lead at M3AAWG.