Machine Learning Library works on Secret Contract


Just out of curiosity, I am researching on if we can use machine learning library in rust under no_std setting and SGX mode.

This time I tried using the following library.

// Built-In Attributes

// Imports
extern crate eng_wasm;
extern crate eng_wasm_derive;
extern crate serde;
extern crate statistical;
extern crate rusty_machine;

extern crate std;

use eng_wasm::*;
use eng_wasm_derive::pub_interface;
use serde::{Serialize, Deserialize};
use std::vec;

use rusty_machine::linalg::Matrix;
use rusty_machine::linalg::Vector;
use rusty_machine::learning::gp::GaussianProcess;
use rusty_machine::learning::gp::ConstMean;
use rusty_machine::learning::toolkit::kernel;
use rusty_machine::learning::SupModel;

// Encrypted state keys
static MILLIONAIRES: &str = "millionaires";

// Structs
#[derive(Serialize, Deserialize)]
pub struct Millionaire {
    address: H160,
    net_worth: U256,

// Public struct Contract which will consist of private and public-facing secret contract functions
pub struct Contract;

// Private functions accessible only by the secret contract
impl Contract {
    fn get_millionaires() -> Vec<Millionaire> {

// Public trait defining public-facing secret contract functions
pub trait ContractInterface{
    fn add_millionaire(address: H160, net_worth: U256);
    fn compute_richest() -> H160;
// Implementation of the public-facing secret contract functions defined in the ContractInterface
// trait implementation for the Contract struct above
impl ContractInterface for Contract {
    fn add_millionaire(address: H160, net_worth: U256) {
        let mut millionaires = Self::get_millionaires();
        millionaires.push(Millionaire {
        write_state!(MILLIONAIRES => millionaires);

    fn compute_richest() -> H160 {
        let v = vec![1.1,2.1,3.1];
        match Self::get_millionaires().iter().max_by_key(|m| m.net_worth) {
            Some(millionaire) => {
                let _tst = statistical::mean(v.as_slice());
                // First we'll get some data.

                // Some example training data.
                let inputs = Matrix::new(3,3,vec![1.,1.,1.,2.,2.,2.,3.,3.,3.]);
                let targets = Vector::new(vec![0.,1.,0.]);

                // Some example test data.
                let test_inputs = Matrix::new(2,3, vec![1.5,1.5,1.5,2.5,2.5,2.5]);

                // Now we'll set up our model.
                // This is close to the most complicated a model in rusty-machine gets!

                // A squared exponential kernel with lengthscale 2, and amplitude 1.
                let ker = kernel::SquaredExp::new(2., 1.);

                // The zero function
                let zero_mean = ConstMean::default();

                // Construct a GP with the specified kernel, mean, and a noise of 0.5.
                let mut gp = GaussianProcess::new(ker, zero_mean, 0.5);

                // Now we can train and predict from the model.

                // Train the model!
                gp.train(&inputs, &targets).unwrap();

                // Predict output from test datae]
                let _outputs = gp.predict(&test_inputs).unwrap();
            None => H160::zero(),

I just added some simple code into the tutorial without exporting the result to the enigma state.
Fortunately, it passed compile, migration, and test code and seems the secret contract can handle machine learning library!!!
I am really feeling Secret Contract will expand the horizon of blockchain world!!!


That’s amazing @KanaGold !

1 Like

Wow, that’s awesome @KanaGold. I totally agree with you on secret contracts expanding the horizons of the blockchain world!

When you’re done, consider sharing with the Enigma team. It would be a great sample repository illustrating a secret contract and Machine Learning use case :slight_smile: .

1 Like

We should look into doing some neural nets or PCA within secret contracts. I know that neural nets for image recognition have small network size and can be added to a secret contract.

Here are some rust sources:

Also we can look into taking an existing library and re-write it in Rust.

1 Like

@KanaGold can you pls tell us a bit more about what the library does? Also what’s your twitter handle?

1 Like

@can the example above is using Gaussian Process Regression for tiny data set. The library covers basic machine learning techniques like K-means, SVM, some NN, and most of the regressions. It seems this library has few dependencies. I am not sure if it is efficiently implemented.

my twitter account is @hereisyourbtg , I basically speak in Japanese on twitter.

Hey @KanaGold – just wanted to let you know a few different teams relied on your ML research here over the past weekend during the Eth Waterloo hacks! Really exciting :). One of those teams ended up winning one of our API prizes!