Request for Source Code: "Machine Learning Training on Encrypted Data with TFHE"

rsynn · September 4, 2024, 7:40am

Hi Zama Team and Community,

I’m a graduate student currently researching the intersection of homomorphic encryption and machine learning. I recently read the paper titled “Machine Learning Training on Encrypted Data with TFHE,” published by Zama, and found it highly relevant to my work.

To further understand and potentially build on the techniques discussed in the paper, I’m kindly requesting access to the source code used in this research. Having the code would greatly assist me in my studies and help me explore this exciting area in more depth.

If the code is available for sharing, I would be extremely grateful. Please let me know the best way to access it, or if there is any additional information needed on my end.

Thank you in advance for your support!

Best regards,
Dancy

jfrery · September 4, 2024, 9:44am

Hello @rsynn,

We published a notebook to showcase how to do the LogisticRegression training you can find it at:

github.com

zama-ai/concrete-ml/blob/main/docs/advanced_examples/LogisticRegressionTraining.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Logistic Regression Training\n",
    "\n",
    "This notebook shows how to train a logistic regression model on encrypted data using stochastic gradient descent (SGD). During this process,\n",
    "the training set remains encrypted at all times and the gradients and loss are encrypted, thus unaccessible by the server performing the training. \n",
    "\n",
    "The result of the encrypted training is a set of encrypted model weights that can only be decrypted by the training set secret-key owner. In Concrete ML the `fit` function encrypts the training data, trains the model producing encrypted weights and then decrypts the weights. The model can then be used in on clear data, or on new encrypted data.\n",
    "\n",
    "Training on encrypted data is especially useful when multiple parties collaborate confidentially, meaning they provide encrypted shares of a training set. \n",
    "\n",
    "In this notebook, a Scikit-Learn model is first trained as the baseline. Next, a Concrete ML model is trained on encrypted data using Fully Homomorphic Encryption."
   ]
  },
  {
   "cell_type": "code",

This file has been truncated. show original

Custom neural network training has not been made public yet but you can look at how we built the SGD training. We basically build a torch module with back propagation in the forward call and return updated parameters. Once we have the torch module, we call the compile_torch_model method that allows us to convert the module to FHE.

github.com

zama-ai/concrete-ml/blob/main/src/concrete/ml/sklearn/_fhe_training_utils.py#L23-L98


      
          class LogisticRegressionTraining(torch.nn.Module):
              """Logistic Regression training module.
          
              We use this torch module to represent the training of a model in order to be able to compile it
              to FHE.
          
              The forward function iterates the SGD over a given certain number of times.
              """
          
              def __init__(self, iterations: int = 1, learning_rate: float = 1.0, fit_bias: bool = True):
                  """Instantiate the model.
          
                  Args:
                      iterations (int): The number of times over which to iterate the SGD during the forward.
                      learning_rate (float): The learning rate.
                      fit_bias (bool): If the bias will be taken into account or not.
                  """
                  super().__init__()
                  self.iterations = iterations
                  self.learning_rate = learning_rate

This file has been truncated. show original

Q_Ma · October 26, 2024, 9:33pm

Hi @jfrery,
I also have a question related to training on encrypted data, especially for custom neural network.
Is the QAT described in here related to this? Can I adjust the model to QAT so it can train on the encrypted data?
Thank you!

Q_Ma · October 26, 2024, 10:16pm

I also noticed in Concrete ML 1.7, which is the latest version, you released that how to fine-tune LLMs on encrypted data, can I leaverage this to other deep learning model when I want to fine-tune the model on the encrypted, domain-specific data

jfrery · October 28, 2024, 7:38am

Hi @Q_Ma,

Is the QAT described in here related to this? Can I adjust the model to QAT so it can train on the encrypted data?

No the QAT you linked is related to train a FHE friendly model for the inference. Training happens in clear here.

I also noticed in Concrete ML 1.7, which is the latest version, you released that how to fine-tune LLMs on encrypted data, can I leaverage this to other deep learning model when I want to fine-tune the model on the encrypted, domain-specific data

Yes! We have a sample code for doing this on a custom MLP: concrete-ml/docs/advanced_examples/LoraMLP.ipynb at main · zama-ai/concrete-ml · GitHub