I am a Ph.D. student in computer science at Columbia University, working with Asaf Cidon, Roxana Geambasu and Mathias Lécuyer.

My research is about privacy-preserving systems, at the intersection of differential privacy, applied cryptography and distributed systems. In particular, I’ve been working on infrastructure systems for differential privacy, such as privacy budget schedulers, protocols for privacy-preserving distributed analytics or large-scale private training of language models.

More generally, I enjoy solving socially meaningful problems by building practical systems that rely on sound theoretical foundations. I also like contributing to open-source projects.

A bit more about me: I did my undergrad in mathematics and computer science at Ecole Polytechnique, in France. My name is pronounced [pjɛʁ to.lo.ɲa]. Feel free to contact me at [email protected].

Infrastructure systems for differential privacy

At Columbia, I am designing and implementing infrastructure systems for differential privacy.

• Privacy budget management and scheduling. I implemented PrivateKube (published at OSDI ‘21), a system to manage privacy budget as a resource (akin to CPU or RAM) on Kubernetes clusters. I later designed more efficient scheduling algorithms, such as DPack (to appear at EuroSys ‘25), to make the most out of this limited resource.
• Private database systems. Turbo (SOSP ‘23) is a caching layer for differential privacy databases that conserves privacy budget while enabling accurate responses to statistical queries. Turbo builds upon private multiplicative weights (PMW), a DP mechanism that is powerful in theory but ineffective in practice, and transforms it into a practical cache that learns from past query results. Turbo is implemented on top of Tumult, a state-of-the-art DP analytics engine, and achieves up to 15.9x budget efficiency on datasets like Covid19 and CitiBike, especially benefiting range-based and streaming queries.
• On-device privacy for web advertising. I presented my most recent project, Cookie Monster, at SOSP ‘24. As major browsers phase out third-party cookies, emerging advertising APIs offer an opportunity to improve web privacy. Cookie Monster enhances existing advertising measurement APIs from major tech companies with more efficient differential privacy (DP) budgeting. By using an individual form of DP, our approach enables more accurate private measurement queries, with additional benefits in terms of user transparency and control. We prototyped Cookie Monster in Chrome, and our design has been incorporated into Mozilla’s draft for standardization at the W3C Private Advertising Technology Working Group.

Check out my lab’s webpage to learn more about these projects.

Past projects

• Large-scale private machine learning. I collaborated with a team at Brookhaven National Laboratory that designs and evaluates large scale privacy-preserving machine learning systems running on top of the U.S. Department of Energy’s supercomputers. We improved support for distributed training in PyTorch’s differential privacy library, Opacus. I worked on similar problems during my internship at Microsoft Research’s Privacy in AI team, where I added differential privacy to mixture-of-experts transformers, a powerful type of large language models.

• Function secret sharing for encrypted deep learning. I am maintaining an open-source library for OpenMined. Sycret is the first Python library for function secret sharing. It relies on an efficient parallelized Rust backend using cryptographic hardware acceleration. We plugged it into PySyft, one of the leading privacy-preserving machine learning frameworks, to build ARIANN: Low-Interaction Privacy-Preserving Deep Learning via Function Secret Sharing (published at PETS ‘22).

• Distributed algorithms. I spent 5 months in 2019 at the Concurrent Systems Research Group of the University of Sydney, where I studied cross-chain protocols and the formal verification of distributed algorithms with Vincent Gramoli and Rob van Glabbeek. In 2020 we worked on a compositional approach to model-check consensus algorithms. Our research appeared at SPAA ‘20, PODC ‘22 and DISC ‘22.