An Open-Source Library for Verifiable Fully Homomorphic Encryption for Trustworthy Machine Learning

Objective

Machine Learning as a Service (MLaaS) is emerging as a cornerstone of the modern digital infrastructure, but its widespread adoption is hampered by critical privacy and integrity concerns. Users must transmit sensitive data to third-party providers, risking privacy breaches and regulatory violations. While Fully Homomorphic Encryption (FHE) offers a powerful solution for privacy by enabling computations on encrypted inputs, it provides no way to verify the correctness of the results. This leaves users vulnerable to errors or malicious manipulation with potentially severe consequences. The VERIFHE project addresses this critical gap in privacy-preserving MLaaS by providing integrity guarantees through verifiable FHE, a cryptographic primitive that, in addition to enabling computations on encrypted inputs, also provides a mechanism to check their correctness. VERIFHE builds upon a recent breakthrough from the ERC-funded PICOCRYPT project—a novel protocol for verifiable FHE—and aims to transform it into a practical, opensource tool. The project’s primary outcomes will be the first-ever open-source software library for verifiable FHE, and a proof-of-concept MLaaS system that, using the library, offers verifiable and privacy-preserving machine learning inference. By delivering the first open-source library for verifiable FHE and demonstrating its feasibility in MLaaS, VERIFHE will accelerate adoption of this technology and pave the way for trustworthy machine learning in industrial sectors where privacy and integrity are paramount.