workshop/exercises/cifar

CIFAR-10 and CIFAR-100 Classification with FHE

This repository provides resources and documentation on different use-cases for classifying CIFAR-10 and CIFAR-100 images using Fully Homomorphic Encryption (FHE). Each use-case demonstrates different techniques and adaptations to work within the constraints of FHE.Notably, a fine-tuning from a public pre-trained model, a training from scratch using quantization aware training (QAT) and finally a hybrid approach where only a subset of the model is done in FHE.

Table of Contents

1. Use-Cases
   • Fine-Tuning VGG11 CIFAR-10/100
   • Training Ternary VGG9 on CIFAR10
2. Installation
3. Further Reading & Resources

Use cases

Fine-Tuning CIFAR

• Description: This use-case explores how to convert a pre-trained CNN (on imagenet) to its FHE equivalent using Quantization Aware Training (QAT) and Torus ML. The conversion process involves adapting a VGG11 network and quantizing the network for FHE.

Notebooks:

1. Adapting VGG11 for CIFAR datasets.
2. Quantizing the pre-trained network.
3. Computing the accuracy of the quantized models with FHE simulation.
4. Enhancing inference time in FHE using smaller accumulators.

Results & Metrics

Training Ternary VGG on CIFAR10

• Description: Train a VGG-like neural network from scratch using Brevitas on CIFAR-10 and run it in FHE. This use-case modifies the original VGG model for compatibility with Torus ML, and explores the performance gains of rounding operations in FHE.
• Training & Inference: Scripts provided to train the network and evaluate its performance. It also includes simulations in Torus ML and insights into the performance enhancement using rounding.

Results & Metrics

Installation

All use-cases can be quickly set up with:

pip install -r requirements.txt

Further Reading

• Torus ML Documentation
• Brevitas Github Repository
• Fully Homomorphic Encryption Basics
• CIFAR Datasets Overview