What is the difference between two bit？ about concrete-ml HOT 2 CLOSED

Thank you， i understand

from concrete-ml.

Hello @maxwellgodv,

Thank you for reaching out to us.

To use Fully Homomorphic Encryption (FHE) with Concrete ML, it is necessary to convert your machine learning model into an FHE-compatible model.

In this use-case, we explain how to convert a custom Torch neural network into its FHE-equivalent.

An FHE-equivalent model means that the model is quantized and the maximum precision of the operation graph is less than 16 bits. So:

• The bit hyper-parameter in QuantVGG11: refers to the quantization bit, an essential hyper-parameter in Concrete-ML, required to quantize the input, weights, activation functions and output. This quantization step is mandatory as FHE operates only over integers with a precision limit of 16 bits. In the provided use-case, we used 5 bits to quantize the model and the inputs.

For custom models, Concrete ml uses Brevitas library for quantization.

• The circuit bit-width :
  Now that your custom neural network is quantized with a precision of 5-bits. You have to check whether the maximum bit-width of your circuit is less than 16 bits. To do so, we use compile_brevitas_qat_model, which under the hood generates an executable operation graph, determines cryptographic parameters and raises an error if the maximum bit-width exceeds 16 bits. In this case, you have to decrease the quantization bit.

Thanks !

from concrete-ml.