zen-blog/content/proof-of-ai.mdx

---
title: "Proof of AI: Verifiable Machine Learning on Chain"
date: "2023-06-26"
description: "How we're bringing cryptographic verification to AI inference, enabling trustless machine learning."
author: "Zach Kelling"
tags: ["Research", "Blockchain", "Verification"]
---


When an AI system makes a prediction, how do you know it actually ran the model it claims? In centralized systems, you trust the operator. Decentralized AI needs cryptographic proof.

Today we introduce Proof of AI (PoAI), a framework for verifiable machine learning inference.

## The Trust Problem

Consider a decentralized AI service:

1. User submits input and payment
2. Compute provider runs inference
3. Provider returns output
4. User receives result

What prevents the provider from:
- Running a cheaper, worse model?
- Returning cached results for new inputs?
- Fabricating outputs entirely?

Traditional solutions require trusted hardware or reputation systems. PoAI provides cryptographic guarantees.

## Proof of AI Overview

PoAI generates succinct proofs that a specific model produced a specific output from a specific input. Verifiers can check proofs efficiently without re-running inference.

### Properties

- **Soundness**: Invalid computations cannot produce valid proofs
- **Completeness**: Valid computations always produce verifiable proofs
- **Succinctness**: Proof size is small relative to computation size
- **Zero-knowledge** (optional): Proofs reveal nothing beyond correctness

### Architecture

```
Input -> [Model Execution] -> Output
              |
              v
         [Circuit]
              |
              v
    [Proof Generation]
              |
              v
          [Proof] -> [Verifier] -> Accept/Reject
```

## Technical Approach

### Model Compilation

Neural networks compile to arithmetic circuits. Each operation becomes constraint equations:

**Matrix multiplication**: $y = Wx + b$ becomes constraints on each element

**Activation functions**: ReLU, GELU approximated by polynomial constraints

**Normalization**: LayerNorm expressed as arithmetic over inputs

Our compiler handles:
- Linear layers
- Attention mechanisms
- Feedforward blocks
- Embedding lookups

### Proof System

We use a combination of techniques:

**SNARKs** for succinct proofs of arithmetic circuits. Proof size is constant regardless of circuit size.

**Folding schemes** to handle the repetitive structure of transformer layers efficiently.

**Lookup arguments** for non-arithmetic operations like embedding tables.

### Optimization

Naive compilation produces impractical circuits. We optimize through:

1. **Quantization**: INT8 models have 8x fewer constraints than FP32
2. **Structured pruning**: Remove entire attention heads, reducing circuit size
3. **Polynomial approximations**: Replace transcendental functions with low-degree polynomials
4. **Batched verification**: Amortize proof costs across multiple inferences

## Performance

### Proof Generation

| Model Size | Parameters | Proof Time | GPU Memory |
|------------|------------|------------|------------|
| Tiny | 25M | 12s | 8GB |
| Small | 110M | 89s | 24GB |
| Medium | 350M | 340s | 48GB |

Proof generation is 100-1000x slower than inference. This is the primary limitation.

### Verification

| Model Size | Verification Time | Proof Size |
|------------|-------------------|------------|
| Tiny | 15ms | 1.2KB |
| Small | 18ms | 1.4KB |
| Medium | 22ms | 1.6KB |

Verification is fast and proof size is nearly constant. On-chain verification is practical.

### Accuracy Impact

Quantization and polynomial approximations affect model accuracy:

| Model | Original Accuracy | PoAI-Compatible | Degradation |
|-------|-------------------|-----------------|-------------|
| Classifier | 94.2% | 93.1% | -1.1% |
| Embeddings | 0.847 (cosine) | 0.831 | -1.9% |
| Generator | 28.3 (perplexity) | 29.1 | +2.8% |

Acceptable for many applications.

## Use Cases

### Decentralized Inference Markets

Users pay for inference, providers compete on price. PoAI ensures providers actually run the claimed model. No reputation bootstrapping needed.

### AI Oracles

Smart contracts need off-chain data. AI models can provide predictions, classifications, or analyses. PoAI makes these oracles trustless.

### Model Verification

When model weights are published, how do you verify they match claimed training? PoAI can prove that specific weights produce specific benchmark results.

### Federated Learning Verification

In federated learning, participants claim to train on local data. PoAI can verify that gradient updates came from actual training, not fabrication.

## Limitations

Current limitations we're working to address:

1. **Proof generation cost**: Large models remain impractical
2. **Model constraints**: Complex architectures (MoE, very deep) are challenging
3. **Floating point**: Native FP support would reduce approximation errors
4. **Recursion**: Autoregressive generation requires sequential proofs

## Roadmap

**Q3 2023**: Release PoAI SDK for small models
**Q4 2023**: Folding scheme improvements for 10x speedup
**Q1 2024**: Support for models up to 1B parameters
**Q2 2024**: Production deployment on Lux Network

## Conclusion

Verifiable AI is essential for decentralized systems. PoAI makes cryptographic verification practical for real models. The overhead is significant but decreasing.

Trust, but verify. Now you can.

---

*Zach Kelling is a co-founder of Zoo Labs Foundation.*