Tancrède Lepoint

AI security & neurosymbolic AI

Secure and reliable AI systems: testing and debugging LLM-integrated applications, securing agentic systems, private training, and neurosymbolic approaches where an LLM interprets and a symbolic engine verifies, so the guarantees never rest on the model alone.

• Delta debugging for LLM-integrated systems LLM applications fail opaquely: no execution paths, no stack traces, non-deterministic outputs. We wrap each input segment in a verifiable semantic marker, have the model cite markers in its reasoning, and use delta debugging to shrink hundreds of segments to the few that cause the failure. ICSE-SEIP 2026.
• Advances and open problems in federated learning The foundational survey of federated learning, covering privacy, robustness, and systems challenges. Foundations and Trends in Machine Learning, 2021.
• Secure single-server aggregation with (poly)logarithmic overhead Efficient secure aggregation protocol for privacy-preserving machine learning. Deployed by Google for federated learning with distributed differential privacy. Published at ACM CCS 2020.

Post-quantum cryptography

Quantum computers will eventually break the cryptographic algorithms that secure today’s internet. I have worked on designing, analyzing, and deploying their replacements.

• CRYSTALS-Kyber and CRYSTALS-Dilithium Lattice-based key encapsulation and digital signature schemes, standardized by NIST as ML-KEM and ML-DSA. Now being deployed worldwide, from TLS to Signal, to protect against quantum threats.
• Improved security proofs in lattice-based cryptography Introduced the use of Rényi divergence as an alternative to statistical distance for lattice-based security reductions, leading to tighter parameters. Best Paper, Asiacrypt 2015.

Privacy-enhancing technologies

Cryptographic systems that let organizations use sensitive data without exposing it: from private information retrieval and anonymous credentials to differential privacy and contact-tracing analytics. Several of these systems have been deployed at scale by Apple and Google.

• Communication–computation trade-offs in PIR Practical homomorphic-encryption-based protocols for private database queries. Deployed by Google for private set membership and by Apple for on-device private information retrieval. Published at Usenix 2021.
• Verified foundations for differential privacy Machine-checked proofs that the core algorithms used in differential privacy implementations are correct. Deployed in AWS Clean Rooms Differential Privacy. Distinguished Artifact, PLDI 2025.
• Exposure Notifications Privacy-preserving Analytics A system for collecting aggregate statistics about COVID-19 exposure notifications without learning anything about individual users. Deployed by Apple and Google.
• Anonymous tokens with private metadata Cryptographic tokens that let servers verify clients without tracking them. Informed the design of Privacy Pass / Trust Tokens, with an IETF/IRTF draft on signature key blinding. Published at Crypto 2020; related ROS cryptanalysis received Best Paper, Eurocrypt 2021.

Open source

fhe.rs A fully homomorphic encryption library in pure Rust, implementing the BFV scheme. More on my GitHub.