Antonio Guimarães

In FHE, one typically considers an idealized model where adversaries are honest-but-curious^?they try to learn the data but follow the protocol. Real-world adversaries may be malicious: they can return incorrect results and use them to mount reaction attacks that recover the data or the secret key. Verifiable FHE lets the client check that the intended computation was performed before decrypting. My work was the first to make this practical for modern FHE schemes, both exact and approximate.

FHE enables multi-party computation (MPC) with communication independent of the function being computed. Two techniques help make it practical: threshold FHE, which distributes the secret key among the parties, and hybrid homomorphic encryption (HHE), which keeps communication concretely low by replacing large FHE ciphertexts with short symmetric ones. I work on both, including threshold decryption that needs no coordination.

Computing over encrypted data may be orders of magnitude slower than in the clear, which has restricted FHE to small-scale applications. I address this at two levels. At the primitive level, I design new algorithms for bootstrapping, which dominates this cost. At the application level, I develop efficient methods for training and inference of machine-learning models over encrypted data. My goal is FHE fast and scalable enough for the models used in practice.

Optimized C library for TFHE: functional and multi-value bootstrapping, blind-rotation unfolding, AVX-512 FFT, and the main TFHE improvements from the literature in one place.

The first open-source library dedicated to verifiable FHE: zero-knowledge proofs, fully homomorphic encryption and related techniques.

RISC-V Assembly Learning Environment: a browser-based RISC-V simulator with virtual devices, built for teaching assembly programming.