Machine Learning in Montpellier, Theory & Practice

Over the past few years, we have seen significant efforts in building trustworthy ML. Many institutions are making privacy and fairness promises about their services. I start the presentation by arguing that institutions might not adhere to their claims of using trustworthy ML across various services intentionally or accidentally due to their interests in maximizing utility and minimizing costs/efforts, resulting in FairWashing /[NeurIPS2022]/ and PrivacyWashing. To address these risks, then, I present Confidential-PROFITT /[Oral ICLR2023] /and Confidential-DPproof /[Spotlight ICLR2024]/ frameworks that enable institutions to directly prove to any interested party through the execution of Zero Knowledge Proof protocols that they train ML models in a fair and privacy-preserving manner, respectively, while protecting the confidentiality of their data and model. [NeurIPS2022] Washing The Unwashable : On The (Im)possibility of Fairwashing Detection, https://openreview.net/pdf?id=3vmKQUctNy [Oral ICLR2023] Confidential-PROFITT: Confidential PROof of FaIr Training of Trees, https://openreview.net/pdf?id=iIfDQVyuFD [Spotlight ICLR2024] Confidential-DPproof: Confidential Proof of Differentially Private Training, https://openreview.net/pdf?id=PQY2v6VtGe