OpenFHE-Tutorials

Installing

1. Install OpenFHE-development
2. Install OpenFHE-python

Caveats:

• the code shown below is highly unoptimized and is meant to be used for educational purposes.

Exercises

There are a total of four exercises:

0. Implementing encrypted inference using the code in exe_encrypted_inference.py. Here, you will load in weights from a pre-trained model (generated from efficient_regression/logreg_reference.ipynb), repeat the weight vector, do the dot-product, and decrypt. See sol_encrypted_inference.py for an example solution.

1. implementing a naive linear regression using the starter code in the naive_regression folder. Work off of exe_lin_reg.py in this top-level folder and see sol_lin_reg.py for one possible solution.

2. Implementing an optimized logistic regression using the starter code in the efficient_regression folder. Work off of the exe_log_reg.py in this top-level folder and see sol_log_reg.py for a possible implementation. You may find it useful to reference the plaintext implementation in logreg_reference.ipynb which shows how it is implemented in raw numpy.

3. Implementing an optimized Nesterov-accelerated gradient logistic regression in the efficient_regression folder. Work off of the exe_nag_log_reg.py and see sol_nag_log_reg.py for a possible implementation. You may find it useful to reference the plaintext implementation in logreg_reference.ipynb which shows how it is implemented in raw numpy.

Tips

Some tips for working with FHE problems:

1. start with a small-ish ring dimension
2. turn off the security setting (via HEStd_NotSet)
3. create a reference numpy implementation
4. Try to do as much as possible in plaintext-space before finally working with ciphertexts
5. ciphertext refreshing speeds up iteration, so start with that for prototyping then move to bootstrapping