Fixing AI models onchain (ZKML)

ZKML

AI technology has emerged as a transformative force in enhancing the efficiency and effectiveness of various systems, notably within the decentralized finance (DeFi) space. However, integrating AI on-chain presents a significant challenge due to the prohibitive computational and financial costs of blockchain transactions. Addressing this challenge, NOYA has pioneered a groundbreaking solution by leveraging an advanced zk-SNARKs (Zero-Knowledge Succinct Non-Interactive Arguments of Knowledge) architecture, known as ZKML (Zero-Knowledge Machine Learning). This innovation, developed in collaboration with Modulus Labs and EZKL, marks a paradigm shift in the on-chain application of AI.

The core essence of ZKML lies in its ability to validate model outputs on-chain trustlessly while at the same time maintaining the privacy of the model's weights. This dual capability ensures that the intellectual property of model architects remains safeguarded, thereby preserving their competitive advantage while not adding any trust assumptions. The underlying principles of this technology can be summarized by:

Privacy Preservation

By hiding the model weights, NOYA ensures that the proprietary aspects of the AI models are not exposed, safeguarding the unique insights and strategies of model developers.

Private models

Enhanced Predictive Capabilities

The framework supports utilizing sophisticated, large-scale AI models off-chain, thereby enabling the deployment of highly accurate predictive analytics in DeFi operations without prohibitive on-chain costs.

Agents using large models for accurate predictions

Trustless Environment

The system fosters a trust-minimized ecosystem by confining on-chain activities to the verification of model outputs through zk-SNARKs. This approach mitigates the risks associated with the manipulation of model predictions.

Only zk-verified movements are executed

Verifiable Performance

Before on-chain deployment, AI models undergo rigorous backtesting. The results of the backtests are published with proofs, ensuring result integrity.

Verifiable performance via zk-proofs of backtested results

Types of agents:

The implementation of ZKML technology within NOYA introduces a variety of AI models, each designed to enhance different aspects of DeFi operations. These models can be broadly classified into four types:

Optimized Private AI Agents:

These models manage the flow of assets in DeFi, ensuring that transactions are efficient and secure. The unique aspect of these models is their ability to keep their strategies hidden, protecting their edge and at the same time handle the movement of on-chain funds.

Private models yet public outputs and proofs

Competing AI Agents:

In certain financial vaults, called competitive vaults, Different private AI models from different model architects are tested against each other. They compete for users' assets by sending consistent zk-proofs of performance.

Better performing models have their results executed

Cooperative Agents:

These AI models are designed to work together, sharing their insights and decisions to create a more comprehensive and effective financial strategy. By combining their strengths, these models can cater to a wider range of needs and preferences. For example, one model can predict volatility while the other predicts liquidity movements.

Specific Agents for Specific Tasks

Restaked Agents:

Some AI models are incredibly powerful and complex, and proving their results on the blockchain can be very costly. Agent architects can restake some assets to vouch for the model's performance. Then, keepers use sophisticated techniques to ensure everything works as it should and the performance is on par with predetermined metrics.

Restaked Agents