N_bits evaluation for Post-Training Quantization

Zama Concrete ML
1

Hi,
I am testing the library to build FHE-compliant models using several fitted scikit-learn models.
I am wondering how to correctly evaluate the n_bits parameter. What aspects should I consider?

For a test, I set n_bits to 14, but all the predictions of my concrete binary problem classifier are wrongly 0.

Thank you!

2

Hello!

Can you share some code? What model and what kind of dataset?

3

Of course.
I published all in this repository: GitHub - CardoFlare/zama-concrete.ml-issue

The dataset is a subset for security reason but is representative of the entire dataset. The model is a XGBoostClassifier for binary classification 0 or 1 .

Problem: the accuracy between the concrete model (fhe enabled or disabled) and the original model (without importing in Concrete) is extremely low.

Note: I recommend using python versione 3.10.15 and creating a virtual environmente to tdo a test in the same development environment.

Thank you in advance!