Astral ZK-ML

Decentralized Machine Learning Protocol adapted to both terrestrial and spatial context. The models and the two smartcontracts utilizing them are written in the ZK-Proovable language Cairo.

✅ ZK-ML tools : Built in Cairo for contracts on starknet-compatible blockchain
    Implemented
        Multi Layer Perceptron (ie. regular Neural Network)
            eg: Star type identification
    Not Implemented Yet
        Convolutional Neural Network
            eg: Crater identification
        Graph Attention Network
            eg: Constellation identification
✅ Decentralized Smart Contracts : Variants of the smart contract :
    Fully Decentralized

        Based on financial incentive (ie. egoistic incentive)
        Terestrial variant adapted to Starknet L2.
    Country Wise Decentralized

        Assumed resistance to Byzantine fault
        Spacial variant adapted to a specific L3 for a planet/satellite owned by multiples countries.
✅ Demonstration Client
    TODO

Country-Wise Decentralized

We assume that regular users are robots owned by multiple countries.
Those countries doesn't want to trust other countries to maintain the algorithms of the robots.
This contract is a proof of concept in the case where the algorithms are machine learning models.

User can call the model to do predictions.

The number of country can increase, if most of the countries agree on the new country.

Each country posseses a list of validators and a level of influence.
Inside a country, most of the validators are assumed loyal.
They doesn't want to degrade the influence level of their country.

Each validator can propose sample for the learning process.

The proposed sample needs to be validated by most validators.
The weight of the vote are determined by the influence of validators's countries.

Validators that are too far from the consensus get a penalty on their countries.

Validator can vote inside their country to replace one of the validators.

Algorithm

We always maintains two "bags" of of input data.

The proposition array. When this array is full, validators can vote on it.
When there is enough votes, the array is destroyed and the survivors are
added to the validated array.

When the validated array has enough data, it launch the training process.

• <p align="center"><img src="../resources/schema_country_wise_decentralized.png" width=870></p>

Miscellaneous

This country-wise version is adapted to spatial context.
Therefore, we assume, a specific Starknet L3 is deployed using madara for the planet (eg: the moon).

Fully Decentralized

Based on financial incentive, adapted on very large network on earth with a lot of users. Implemented on Starknet.