Human In Picture (HIP) using Concrete ML

Objective

• Create a CNN using Concrete ML to identify if there is a human in a picture.
• Identify the performance limits of a model for this problem using Concrete ML.
• Create benchmarks for different sizes of the input.

Requirements

Kaggle account

• You will need to have a Kaggle account in order to download the dataset, we recommend using Google Login.
• After creating an account you will need to download a kaggle.json file as the API key.
• You can find that file by going to Your Profile and scrolling down to the API section.
• Then create a new token and you will download the kaggle.json.
• Create a new hidden dictory called .kaggle in the home directory.
• Move kaggle.json to .kaggle

Python version

• Required Python version: 3.10 < 3.11
• This project uses Poetry. If you don't already have Poetry installed, make deps will install it for you.
  • Make sure to have installed curl before running the command.

Installation

Dependencies

• To install dependencies you will need to run:

make deps

Dataset

• To download the dataset you will need to run:

make data

Running the project

Single run

• To run the project only once you'll need to run:

make run

Benchmarks

• To run the benchmarks you'll need to run:

make benchmark

• This will run the benchmarks for the following input sizes:
  • 32x32
  • 64x64
  • 96x96
  • 128x128
• Brace yourself, this will take a while.

Webapp

• The webapp is a proof of concept heavily inspired by Zama's encrypted image filtering example found on hugging face.
• To run the webapp locally the following command is needed:

make webapp

• A more detailed explanation can be found in the webapps README file.

References

• Introduction to homomorphic encryption
• Zama Whitepaper.
• Homomorphic Encryption 101.
• Fully Homomorphic Encryption Using Ideal Lattices.
• Concrete ML from Zama.
• Encrypted Image Filtering Using Homomorphic Encryption Blog Post
• MNIST Concrete ML example.
• Convolutional Neural Network advanced example.
• human-vs-non-human dataset used for training.

---