News Archive

June 2026 Our article Learning with Differentially Private Sliced Wasserstein Gradients just got accepted for publication in the journal Transactions on Machine Learning Research. Huge thanks to my coauthors David Rodríguez-Vítores and Jean-Michel Loubes, and to the reviewers for their suggestions.

Try it: training an encoder on private data with differentially private sliced Wasserstein gradients.

Along the dashed direction: encoded data (top) matched in sorted order with the prior (bottom)

Privacy budget Private data size

private trainingsame run without privacy

Solid dots: the private data after encoding; faint dots: the public prior N(0, I). At each step, the gradient of the sliced Wasserstein distance is computed along a few random directions, where optimal transport simply matches sorted values. The data and the encoded points are clipped, and Gaussian noise calibrated to the gradient's sensitivity is added before each Adam update, so that the whole training is (ε, 10⁻⁵)-differentially private. Our main result bounds this sensitivity by O(1/n): with more data, privacy costs less noise, and the private encoder gets close to the non-private one.
June 2026 It was great to talk about change detection in LLMs at TSE! Thanks to the organizers Adrien Blanchet and Julien Chhor. Slides available in .pdf and .key formats.
June 2026 New preprint: Minimax Private Estimation of Smooth Optimal-Transport Maps (Clément Lalanne, David Rodríguez-Vítores, Franck Iutzeler, Jean-Michel Loubes).
May 2026 Our article Token-Efficient Change Detection in LLM APIs (Timothée Chauvin, Clément Lalanne, Erwan Le Merrer, Jean-Michel Loubes, François Taïani, Gilles Tredan) just got accepted to ICML 2026! See you in Seoul :)
February 2026 New preprint: Token-Efficient Change Detection in LLM APIs (Timothée Chauvin, Clément Lalanne, Erwan Le Merrer, Jean-Michel Loubes, François Taïani, Gilles Tredan). We show that LLM change detection exhibits a sharp low-temperature phase transition, and leverage this phenomenon to build a token-efficient method that reduces API costs by up to 30x.
January 2026 I was delighted to be invited to speak in the session "Theory, Data, Models: What Foundations for AI?" during the ANITI Days 2026. Many thanks to Edouard Pauwels and Nicholas Asher for their invitation. It was a little intimidating to speak alongside David Donoho, Emiliano Lorini, and Rufin VanRullen, but everything went smoothly in the end. For anyone interested, the slides are available in .pdf and .key formats.
December 2025 Had a great time at EurIPS 2025 in Copenhagen, Denmark. I was happy to attend and connect with the community during the workshop Privacy-Preserving Machine Learning: Towards Tight Guarantee. It was a stimulating and inspiring week, with many insightful discussions and high-quality talks. Many thanks to the organizers and participants for making this event such a success!
October 2025 I had the pleasure of attending the workshop Probabilistic Metrics and Trimming Techniques in Statistics, organized in honour of Juan A. Cuesta-Albertos and Carlos Matrán. Many thanks to the organizers for this very nice and stimulating meeting. It was a pleasure to see a workshop dedicated to two researchers whose work has played an important role in both probability theory and statistics, surrounded by such a nice and welcoming community!
July 2025 Had a great time at ICML 2025 in Vancouver, Canada, where I presented a poster for our article On the Private Estimation of Smooth Transport Maps. It was a fun and inspiring week; I really enjoyed the talks and connecting with so many brilliant people. Hats off to the organizers and presenters!

Try it: privately selecting a transport map among N candidate potentials, in one dimension.

Candidate maps, shown as the displacement T(x) − x

true map private choice earlier draws choice without noise candidates

Semi-dual scores minus the best one (lower is better)

score on the data score + Laplace noise smallest noisy score

Privacy budget Sample size Number of candidates

Our estimator scores each candidate Brenier potential f by the semi-dual objective on the private data, (1/n) Σ f(X_i) + (1/n) Σ f*(Y_i), which is smallest at the true potential, and returns the candidate with the smallest score after adding Laplace noise of scale 4C/(nε): this "report noisy argmin" is ε-differentially private. The estimated map is the gradient of the chosen potential. As in the experiment of our paper, the candidates are random Gaussian attraction/repulsion potentials, and scores are computed on a grid and clipped at C = 0.25. With little data or privacy budget, the noise swamps the differences between scores and the choice is nearly random; with more candidates, a better one exists, but it takes more data to find it privately.
June 2025 Had a great time at JdS 2025 in Marseille, France. Such a great community!
May 2025 The preprint for our article Learning with Differentially Private (Sliced) Wasserstein Gradients (David Rodríguez-Vítores, Clément Lalanne, Jean-Michel Loubes) just got updated, including some interesting new experiments on private sliced-Wasserstein autoencoders!
May 2025 Our article On the Private Estimation of Smooth Transport Maps (Clément Lalanne, Franck Iutzeler, Jean-Michel Loubes, Julien Chhor) just got accepted to ICML 2025! The camera-ready version of the article should soon be available, including some nice additions following the discussions with the reviewers. See you in Vancouver :)
January 2025 The preprint for our new paper On the Private Estimation of Smooth Transport Maps (Clément Lalanne, Franck Iutzeler, Jean-Michel Loubes, Julien Chhor) is now available. Check it out!
January 2025 The preprint for our new paper Learning with Differentially Private (Sliced) Wasserstein Gradients (David Rodríguez-Vítores, Clément Lalanne, Jean-Michel Loubes) is now available. Read it now!
September 2024 I am thrilled to announce that I started a new position as a Maître de Conférences (the French equivalent of Assistant Professor) at Université Toulouse 3 Paul Sabatier in Applied Mathematics. I will be pursuing my research within the Statistics and Optimization team of the Institut de Mathématiques de Toulouse.
July 2024 What a week at ICML 2024 in Vienna, Austria, where I presented a poster for our article Privately Learning Smooth Distributions on the Hypercube by Projections. Congratulations to the researchers for the high level of the presentations and to the organization team.
May 2024 Our article Privately Learning Smooth Distributions on the Hypercube by Projections just got accepted to the 41st International Conference on Machine Learning. Looking forward to presenting our work in Vienna. Huge thanks to my coauthor Sébastien Gadat, and to the reviewers for the useful inputs and suggestions.
October 2023 I have just started a new position as a postdoctoral researcher in Mathematics with Jérôme Bolte and Sébastien Gadat at Toulouse School of Economics, Université Toulouse 1 Capitole!
October 2023 It is with great pleasure that I announce that on October 4, I successfully defended my PhD thesis in computer science titled "On the tradeoffs of statistical learning with privacy" at ENS Lyon. Huge thanks to everyone who made this possible, and in particular to my advisors Aurélien Garivier and Rémi Gribonval, to the referees Aurélien Bellet and Béatrice Laurent-Bonneau, and to the president of the jury Élisa Fromont.
September 2023 Just got back from GRETSI 2023 in beautiful Grenoble, France. Really interesting to attend this human-size signal-processing conference, and to see presentations outside of my usual topics of interest. Congrats to everyone who presented and to the organization team. I am still blown away by the amazing performance of the theater troupe Le Gravillon for their play "Point de Bascule" (social activity during the conference)!
August 2023 Our article About the Cost of Central Privacy in Density Estimation just got accepted for publication in the journal Transactions on Machine Learning Research. Huge thanks to my coauthors Aurélien Garivier and Rémi Gribonval.
July 2023 Just got back from ICML 2023 in beautiful Honolulu, HI, USA, where I presented a poster for our article Private Statistical Estimation of Many Quantiles. From both a scientific and a human point of view, it was amazing to meet so many passionate people. Congratulations to the researchers for the high level of the presentations and to the organization team.
July 2023 I am glad to announce that I am a laureate of the 2023 "Agrégation Externe de Mathématiques" (as an independent candidate during my PhD).
June 2023 Our article Private Quantiles Estimation in the Presence of Atoms just got accepted for publication in the journal Information and Inference. You will soon be able to read it with a new template and some important updates thanks to the help from the reviewers. Also, huge thanks to my coauthors Clément Gastaud and Nicolas Grislain from Sarus Technologies, Aurélien Garivier and Rémi Gribonval.
June 2023 Our article/experiment report Sparsity in neural networks can improve their privacy just got accepted to the French XXIXème Colloque Francophone de Traitement du Signal et des Images (GRETSI 2023 Conference). Huge thanks to the reviewers for their suggestions. Also, huge thanks to my coauthors Antoine Gonon, Léon Zheng, Quoc-Tung Le, Guillaume Lauga and Can Pouliquen. We are currently working on a scaled and extended version of this article.
June 2023 Huge thanks to Dataiku for hosting the Privacy Preserving Machine Learning workshop on June 14, 2023 in Paris, and to everyone who contributed to the organization. Congratulations to everyone who presented for the high quality of the presentations!
April 2023 Our article Private Statistical Estimation of Many Quantiles just got accepted to the 40th International Conference on Machine Learning. Looking forward to presenting our work in Honolulu. Huge thanks to my coauthors Aurélien Garivier and Rémi Gribonval, and to the reviewers for the useful inputs and suggestions.
April 2023 Our article On the Statistical Complexity of Estimation and Testing under Privacy Constraints just got accepted for publication in the journal Transactions on Machine Learning Research. You will soon be able to read it with a new template and some important updates thanks to the positive inputs from the reviewers. Also, huge thanks to my coauthors Aurélien Garivier and Rémi Gribonval.
November 2022 Huge thanks to Meta for hosting the Privacy Preserving Machine Learning workshop on November 9, 2022 in Paris, and to everyone who contributed to the organization. Congratulations to everyone who presented for the high quality of the presentations!
June 2022 Huge thanks to Sarus Technologies for hosting the Privacy Preserving Machine Learning workshop on June 8, 2022 in Paris, and to everyone who contributed to the organization. Congratulations to everyone who presented for the high quality of the presentations!