Thesis of Gabriel Loiseau

Task-Oriented Privacy and Utility Preserving Text Anonymization

Text anonymization enables sensitive documents to be reused without exposing the people they involve. In free-form text, however, privacy is not confined to explicit identifiers: it is distributed across semantic, contextual, and stylistic cues whose removal can also destroy useful information. A task-oriented view offers a way to make this tension operational. On the privacy side, privacy-evaluation tasks turn a threat model into a measurable objective, for example by testing what a specified attacker may infer from the transformed text. On the utility side, downstream tasks specify which information a particular use requires, relaxing the overly demanding constraint that an anonymization method preserve the text as a whole. Because privacy and utility requirements vary with the threat model, domain, and intended use, no transformation is universally optimal. This thesis therefore studies anonymization as a task-oriented privacy–utility optimization problem. We ask how this setting should be formulated and measured, and how anonymization systems can adapt to different requirements. We first address evaluation. We introduce a modular framework that pairs a text-to-text transformation with a privacy-evaluation task grounded in an explicit threat model and with a downstream utility task, while retaining generic measures of the transformation itself. Experiments across heterogeneous settings show that anonymization effects depend on the task being considered and that generic similarity measures provide incomplete and inconsistent proxies for operational utility. We then study perceived sensitivity as an alternative privacy-evaluation task. By distilling judgments from a large teacher model into compact local encoders, we obtain a scalable evaluator that approximates aggregate human judgments and complements privacy evaluation through explicit attack tasks. We then turn from evaluation to methods that optimize text transformations against explicit privacy and utility objectives. In authorship obfuscation, we formulate generative rewriting as multi-objective policy optimization guided by explicit privacy and utility rewards. The resulting models reduce authorship attribution while preserving task-relevant information, and their privacy gains persist against attackers retrained on anonymized text. We next move adaptation from model weights to instructions. Reflective prompt optimization learns task- and model-specific anonymization instructions for fixed language models and retains a candidate population whose post-hoc evaluation reveals multiple non-dominated privacy–utility operating points. Together, these contributions make privacy–utility trade-offs explicit enough to evaluate and optimize. Rather than prescribing a universal transformation, they support the selection of documented operating points suited to a concrete threat model, domain, and downstream use.

Jury

- M. Marc TOMMASI Professeur des universités Université de Lille Directeur de thèse, - M. Damien SILEO Chargé de recherche Inria Lille Co-encadrant de thèse, - Mme Aurélie NéVéOL Directrice de recherche LISN CNRS Rapporteure, - M. Ivan HABERNAL Professeur Ruhr University Bochum Rapporteur, - Mme Claire GARDENT Directrice de recherche LORIA Examinatrice, - M. Pierre LISON Professeur associé Norsk Regnesentral Examinateur, - M. Maxime MEYER Hornetsecurity Invité, - M. Damien RIQUET Hornetsecurity Invité.

Thesis of the team MAGNET defended on 02/11/2026