2026
BA-FedSHAP: A reproducible toolkit for auditing background-induced attribution drift
SSRN
Preprint on how the choice of background distribution alters federated attribution results, and what that does to an audit.
PublicationsAuthor
Domaine de recherche
Des institutions qui ne peuvent pas mettre leurs données sensibles en commun peuvent tout de même apprendre ensemble, à condition que la gouvernance soit construite avec la méthode.
Que faut-il établir, techniquement et institutionnellement, pour qu'une analyse fédérée sur des données sensibles soit digne de confiance et pas seulement décentralisée ?
Federated learning is often presented as a privacy guarantee. It is better understood as a change in where the risk sits. Moving computation to the data removes one exposure and introduces others: attribution that shifts with modelling choices, updates that leak, and a governance question about who is accountable when no single party holds the dataset.
I work on both halves of that, the technical behaviour of federated methods under audit, and the institutional arrangements that make a federation legible to a regulator or a ministry.
Productions
2026
SSRN
Preprint on how the choice of background distribution alters federated attribution results, and what that does to an audit.
PublicationsAuthor
2026
Measures how much a federated explanation depends on an arbitrary modelling choice.
Research softwareAuthor
2025
UNICEF Innocenti, Office of Strategy and Evidence
Soutient que les États peuvent exploiter des données éducatives sensibles sans les centraliser, et précise la gouvernance que cela suppose.
PolicyAuthor