Tutorial Overview

The rise of Machine-Learning-as-a-Service (MLaaS) has transformed how machine learning (ML) models are developed and deployed in cloud environments. However, this growth has been accompanied by significant security challenges such as data leakage and the looming threat of quantum computing, which can compromise traditional cryptographic systems. Fully Homomorphic Encryption (FHE), particularly the Cheon–Kim–Kim–Song (CKKS) scheme, has emerged as a powerful post-quantum approach that enables computation directly on encrypted data. Despite its potential, the computational cost of CKKS—up to five orders of magnitude slower than plaintext computation—remains a key barrier to adoption.

To address this, we introduce FIDESlib, the first open-source library that implements the complete CKKS functionality on GPUs and multi-GPUs, including bootstrapping. Built on top of OpenFHE, FIDESlib leverages GPU acceleration to significantly improve performance and accessibility for privacy-preserving MLaaS workloads. The library includes comprehensive testing and benchmarking tools (GoogleTest, Google Benchmark) to ensure robustness and reproducibility.

Objectives

Through this tutorial, participants will learn the theoretical and practical aspects of CKKS-based FHE and how FIDESlib enables efficient, scalable, and open GPU-based private computation in the cloud.

The major takeaways for participants will include:

• Understand the fundamentals of FHE and the CKKS scheme for privacy-preserving computation in MLaaS.

• Identify computational challenges and performance bottlenecks of CKKS on CPU-based systems.

• Explore GPU acceleration as a practical means to enable efficient private computation in cloud environments.

• Gain hands-on experience using FIDESlib to implement and optimize CKKS-based workloads on GPU platforms.

• Contribute to the open-source development and research community around GPU-based FHE.

Target Audience

Graduate students, researchers, and educators interested in secure computation, GPU programming, and cryptographic systems for MLaaS. A basic understanding of linear algebra, modular arithmetic, and parallel programming (CUDA, OpenCL, or HIP) is recommended.

Hardware and Software Requirements

• Participants should bring laptops equipped with NVIDIA or AMD GPUs. 

• Recommended system: Ubuntu OS with Docker preinstalled and configured to have access to GPU devices.

• A Docker image with FIDESlib preinstalled will be provided to simplify setup

FIDESlib GitHub repo

Recommended Reading before the Tutorial

The Beginner’s Textbook for Fully Homomorphic Encryption by Ronny Ko on Sept 8, 2025

FIDESlib publication and its extended version

Citation

If you use FIDESlib on your research, please cite our ISPASS paper: