Circom Circuits Library for Machine Learning over the Vesta Curve Prime

Circom circuits for ML, forked from Socathie's repo. Designed for the Vesta Curve, but will work with BN-254 as long as all numbers are < p // 2 where p is the BN-254 prime.

Run npm run test to test all circomlib-ml circuit templates.

_Disclaimer: This package is not affiliated with circom, circomlib, or iden3._

Description

- This repository contains a library of circuit templates.
- You can read more about the circom language in the circom documentation webpage.

Organisation

This respository contains 2 main folders:
- circuits: all circom files containing ML templates and relevant util templates
``bash
circuits/
├── ArgMax.circom
├── AveragePooling2D.circom
├── BatchNormalization2D.circom
├── Conv1D.circom
├── Conv2D.circom
├── Dense.circom
├── Flatten2D.circom
├── MaxPooling2D.circom
├── Poly.circom
├── ReLU.circom
├── SumPooling2D.circom
├── circomlib
│ ├── aliascheck.circom
│ ├── babyjub.circom
│ ├── binsum.circom
│ ├── bitify.circom
│ ├── comparators.circom
│ ├── compconstant.circom
│ ├── escalarmulany.circom
│ ├── escalarmulfix.circom
│ ├── mimc.circom
│ ├── montgomery.circom
│ ├── mux3.circom
│ ├── sign.circom
│ └── switcher.circom
├── circomlib-matrix
│ ├── matElemMul.circom
│ ├── matElemSum.circom
│ └── matMul.circom
├── crypto
│ ├── ecdh.circom
│ ├── encrypt.circom
│ └── publickey_derivation.circom
└── util.circom
`
- test: containing all test circuits and unit tests
`bash
test/
├── AveragePooling2D.js
├── BatchNormalization.js
├── Conv1D.js
├── Conv2D.js
├── Dense.js
├── Flatten2D.js
├── IsNegative.js
├── IsPositive.js
├── Max.js
├── MaxPooling2D.js
├── ReLU.js
├── SumPooling2D.js
├── circuits
│ ├── AveragePooling2D_stride_test.circom
│ ├── AveragePooling2D_test.circom
│ ├── BatchNormalization_test.circom
│ ├── Conv1D_test.circom
│ ├── Conv2D_stride_test.circom
│ ├── Conv2D_test.circom
│ ├── Dense_test.circom
│ ├── Flatten2D_test.circom
│ ├── IsNegative_test.circom
│ ├── IsPositive_test.circom
│ ├── MaxPooling2D_stride_test.circom
│ ├── MaxPooling2D_test.circom
│ ├── Max_test.circom
│ ├── ReLU_test.circom
│ ├── SumPooling2D_stride_test.circom
│ ├── SumPooling2D_test.circom
│ ├── decryptMultiple_test.circom
│ ├── decrypt_test.circom
│ ├── ecdh_test.circom
│ ├── encryptDecrypt_test.circom
│ ├── encryptMultiple_test.circom
│ ├── encrypt_test.circom
│ ├── encrypted_mnist_latest_test.circom
│ ├── mnist_convnet_test.circom
│ ├── mnist_latest_precision_test.circom
│ ├── mnist_latest_test.circom
│ ├── mnist_poly_test.circom
│ ├── mnist_test.circom
│ ├── model1_test.circom
│ └── publicKey_test.circom
├── encryption.js
├── mnist.js
├── mnist_convnet.js
├── mnist_latest.js
├── mnist_latest_precision.js
├── mnist_poly.js
└── model1.js
`

Template descriptions:


- ArgMax.circom

ArgMax(n) takes an input array of length n and returns an output signal correponds to the 0-based position index of the maximum element in the input.

- AveragePooling2D.circom

AveragePooling2D (nRows, nCols, nChannels, poolSize, strides, scaledInvPoolSize) takes an nRow-by-nCol-by-nChannels input array and performs average pooling on patches of poolSize-by-poolSize with step size strides. scaleInvPoolSize must be computed separately and supplied to the template. For example, a poolSize of 2, should have an scale factor of 1/(2*2) = 0.25 --> 25.

- BatchNormalization2D.circom

BatchNormalization2D(nRows, nCols, nChannels) performs Batch Normalization on a 2D input array. To avoid division, parameters in the layer is parametrized into a multiplying factor a and constant addition b, i.e. ax+b, where

`python
a = gamma/(moving_var+epsilon)**.5
b = beta-gammamoving_mean/(moving_var+epsilon)*.5
`

- Conv1D.circom

1D version of Conv2D

- Conv2D.circom

Conv2D (nRows, nCols, nChannels, nFilters, kernelSize, strides) performs 2D convolution on an nRow-by-nCol-by-nChannels input array with filters of kernelSize-by-kernelSize and step size of strides. Currently it works only for "valid" padding setting.

- Dense.circom

Dense (nInputs, nOutputs) performs simple matrix multiplication on the 1D input array of length nInputs and weights then adds biases to the output.

- Flatten2D.circom

Flatten2D (nRows, nCols, nChannels) flattens an nRow-by-nCol-by-nChannels input array.

- MaxPooling2D.circom

MaxPooling2D (nRows, nCols, nChannels, poolSize, strides)

- Poly.circom

Inspired by Ali, R. E., So, J., & Avestimehr, A. S. (2020), the Poly() template has been addded as a template to implement f(x)=x2+x as an alternative activation layer to ReLU. The non-polynomial nature of ReLU activation results in a large number of constraints per layer. By replacing ReLU with the polynomial activation f(n,x)=x2+n*x, the number of constraints drastically decrease with a slight performance tradeoff. A parameter n is required when declaring the component to adjust for the scaling of floating-point weights and biases into integers. See below for more information.

- ReLU.circom

ReLU() takes a single input and performs ReLU activation, i.e. max(0,x). This computation is much more expensive than Poly(). It's recommended to adapt your activation layers into polynomial activation to reduce the size of the final circuit.

- SumPooling2D.circom

Essentially AveragePooling2D with a constant scaling of poolSize*poolSize. This is preferred in circom to preserve precision and reduce computation.

Weights and biases scaling:


- Circom only accepts integers as signals, but Tensorflow weights and biases are floating-point numbers.
- In order to simulate a neural network in Circom, weights must be scaled up by 10**m times. The larger m is, the higher the precision.
- Subsequently, biases (if any) must be scaled up by 10**2m times or even more to maintain the correct output of the network.

An example is provided below.

Scaling example: mnist_poly


In models/mnist_poly.ipynb, a sample model of Conv2d-Poly-Dense layers was trained on the MNIST dataset. After training, the weights and biases must be properly scaled before inputting into the circuit:
- Pixel values ranged from 0 to 255. In order for the polynomial activation approximation to work, these input values were scaled to 0.000 to 0.255 during model training. But the original integer values were scaled by 10**6 times as input to the circuit
- Overall scaled by 10**9 times
- Weights in the Conv2d layer were scaled by 109 times for higher precision. Subsequently, biases in the same layer must be scaled by (109)*(109)=1018 times.
- The linear term in the polynomial activation layer would also need to be adjusted by 1018 times in order to match the scaling of the quadratic term. Hence we performed the acitvation with f(x)=x2+(10*18)x.
- Weights in the Dense layer were scaled by 10**9 time for precision again.
- Biases in the Dense layer had been omitted for simplcity, since ArgMax layer is not affected by the biases. However, if the biases were to be included (for example in a deeper network as an intermediate layer), they would have to be scaled by (109)5=10**45 times to adjust correctly.

We can easily see that a deeper network would have to sacrifice precision, due to the limitation that Circom works under a finite field of modulo p which is around 254 bits. As log(2254)~76, we need to make sure total scaling do not aggregate to exceed 1076 (or even less) times. On average, a network with l layers should be scaled by less than or equal to 10**(76//l)` times.