Stepbystep guide
This guide demonstrates how to convert a PyTorch neural network into a Fully Homomorphic Encryption (FHE)friendly, quantized version. It focuses on Quantization Aware Training (QAT) using a simple network on a synthetic dataset. This guide is based on a notebook tutorial, from which some code blocks are documented.
Quantization
In general, quantization can be carried out in two different ways:
During the training phase with Quantization Aware Training (QAT)
After the training phase with Post Training Quantization (PTQ).
For FHEfriendly neural networks, QAT is the best method to achieve optimal accuracy under FHE constraints. This technique reduces weights and activations to very low bitwidths (for example, 23 bits). When combined with pruning, QAT helps keep low accumulator bitwidths.
Concrete ML uses the thirdparty library Brevitas to perform QAT for PyTorch neural networks, but options exist for other frameworks such as Keras/Tensorflow. Concrete ML provides several demos and tutorials that use Brevitas , including the CIFAR classification tutorial. For a more formal description of the usage of Brevitas to build FHEcompatible neural networks, please see the Brevitas usage reference.
For a formal explanation of the mechanisms that enable FHEcompatible neural networks, please see the the following paper.
Deep Neural Networks for Encrypted Inference with TFHE, 7th International Symposium, CSCML 2023
Baseline PyTorch model
In PyTorch, using standard layers, a Fully Connected Neural Network (FCNN) would look like this:
Similarly to the one above, the notebook tutorial shows how to train a FCNN on a synthetic 2D dataset with a checkerboard grid pattern of 100 x 100 points. The data is split into 9500 training and 500 test samples.
Once trained, you can import this PyTorch network using the compile_torch_model
function, which uses simple PTQ.
The network was trained using different numbers of neurons in the hidden layers, and quantized using 3bits weights and activations. The mean accumulator size, shown below, is measured as the mean over 10 runs of the experiment. An accumulator size of 6.6 means that 4 times out of 10, the accumulator was 6 bits, while 6 times it was 7 bits.
neurons  10  30  100 

fp32 accuracy  68.70%  83.32%  88.06% 
3bit accuracy  56.44%  55.54%  56.50% 
mean accumulator size  6.6  6.9  7.4 
This shows that the fp32 accuracy and accumulator size increases with the number of hidden neurons, while the 3bits accuracy remains low regardless of the number of neurons. Although all configurations tested were FHEcompatible (accumulator < 16 bits), it is often preferable to have a lower accumulator size to speed up inference time.
Accumulator size is determined by Concrete as the maximum bitwidth encountered anywhere in the encrypted circuit.
Quantization Aware Training (QAT)
Using QAT with Brevitas is the best way to guarantee a good accuracy for Concrete ML compatible neural networks.
Brevitas provides quantized versions of almost all PyTorch layers. For example, Linear
layer becomes QuantLinear
, and ReLU
layer becomes QuantReLU
. Brevitas also offers additional quantization parameters, such as:
bit_width
: precision quantization bits for activationsact_quant
: quantization protocol for the activationsweight_bit_width
: precision quantization bits for weightsweight_quant
: quantization protocol for the weights
To use FHE, the network must be quantized from end to end. With the Brevitas QuantIdentity
layer, you can quantize the input by placing it at the network's entry point. Moreover, you can combine PyTorch and Brevitas layers, as long as a QuantIdentity
layer follows the PyTorch layer. The following table lists the replacements needed to convert a PyTorch neural network for Concrete ML compatibility.
PyTorch fp32 layer  Concrete ML model with PyTorch/Brevitas 









Some PyTorch operators (from the PyTorch functional API), require a brevitas.quant.QuantIdentity
to be applied on their inputs.
PyTorch ops that require QuantIdentity 





The QAT import tool in Concrete ML is a work in progress. While it has been tested with some networks built with Brevitas, it is possible to use other tools to obtain QAT networks.
With Brevitas, the network above becomes:
In the network above, biases are used for linear layers but are not quantized ("bias": True, "bias_quant": None
). The addition of the bias is a univariate operation and is fused into the activation function.
Training this network with pruning (see below) using 30 out of 100 total nonzero neurons gives good accuracy while keeping the accumulator size low.
Nonzero neurons  30 

3bit accuracy brevitas  95.4% 
3bit accuracy in Concrete ML  95.4% 
Accumulator size  7 
The PyTorch QAT training loop is the same as the standard floating point training loop, but hyperparameters such as learning rate might need to be adjusted.
QAT is somewhat slower than normal training. QAT introduces quantization during both the forward and backward passes. The quantization process is inefficient on GPUs due to its low computational intensity is low relative to data transfer time.
Pruning using Torch
Considering that FHE only works with limited integer precision, there is a risk of overflowing in the accumulator, which will make Concrete ML raise an error.
To understand how to overcome this limitation, consider a scenario where 2 bits are used for weights and layer inputs/outputs. The Linear
layer computes a dot product between weights and inputs $y = \sum_i w_i x_i$. With 2 bits, no overflow can occur during the computation of the Linear
layer as long the number of neurons does not exceed 14, as in the sum of 14 products of 2bits numbers does not exceed 7 bits.
By default, Concrete ML uses symmetric quantization for model weights, with values in the interval $\left[2^{n_{bits}1}, 2^{n_{bits}1}1\right]$. For example, for $n_{bits}=2$ the possible values are $[2, 1, 0, 1]$; for $n_{bits}=3$, the values can be $[4,3,2,1,0,1,2,3]$.
In a typical setting, the weights will not all have the maximum or minimum values (such as $2^{n_{bits}1}$). Weights typically have a normal distribution around 0, which is one of the motivating factors for their symmetric quantization. A symmetric distribution and many zerovalued weights are desirable because opposite sign weights can cancel each other out and zero weights do not increase the accumulator size.
This fact can be leveraged to train a network with more neurons, while not overflowing the accumulator, using a technique called pruning where the developer can impose a number of zerovalued weights. Torch provides support for pruning out of the box.
The following code shows how to use pruning in the previous example:
Results with PrunedQuantNet
, a pruned version of the QuantSimpleNet
with 100 neurons on the hidden layers, are given below, showing a mean accumulator size measured over 10 runs of the experiment:
Nonzero neurons  10  30 

3bit accuracy  82.50%  88.06% 
Mean accumulator size  6.6  6.8 
This shows that the fp32 accuracy has been improved while maintaining constant mean accumulator size.
When pruning a larger neural network during training, it is easier to obtain a low bitwidth accumulator while maintaining better final accuracy. Thus, pruning is more robust than training a similar, smaller network.
Last updated