ZKFair

ZKFair is a trustless platform for proving ML model fairness using Zero-Knowledge Proofs.

• Privacy-Preserving: Prove fairness without exposing training data or model weights
• Trustless: On-chain commitments + ZK proofs eliminate trusted third parties
• Continuous: Runtime auditing of inference queries, not just training
• Economic Security: Stake-based incentives with slashing for misbehavior

---

How It Works

1. Certify

Provider registers model on-chain with cryptographic commitments. ZK proof verifies training fairness.

2. Serve

Provider runs inference, issues signed receipts. Queries batched into Merkle trees on-chain.

3. Audit

Anyone can challenge batches. Provider must respond with ZK fairness proof within 24h.

4. Detect

Users verify receipts against blockchain. Disputes slash malicious providers.

---

Get Started

Quick Start End-to-end walkthrough for providers, users, and challengers.

Architecture System design, components, and security model.

SDK Reference TypeScript SDK for ProviderSDK and BrowserSDK.

ZK Circuits Noir circuits for training and audit proofs.

---

Core Components

Component	Description
CLI	Register models, generate proofs, submit certifications
Server	Inference API with batching, receipts, audit handling
Web App	User interface with Sentinel fraud detection
SDK	ProviderSDK (server) and BrowserSDK (client)
Contracts	Model registry, batching, disputes, slashing
ZK Circuits	Training certification + fairness audit proofs

---

Protocol Phases

Phase 1: Certification

Provider registers model with cryptographic commitments:

# Hash weights + build dataset Merkle tree
zkfair commit --dir ./my-model

# Generate ZK training proof + submit
zkfair proof generate-and-submit --weights-hash 0x...

Phase 2: Serving

Provider runs inference and batches queries:

// POST /predict
const receipt = await provider.createSignedReceipt({
  seqNum, modelId, features, prediction, timestamp
});

// Auto-batch when thresholds met
await provider.createBatchIfNeeded();

Phase 3: Auditing

Anyone can challenge batches:

// Request audit (stakes 0.00005 ETH)
requestAudit(batchId);

// Provider auto-responds with ZK proof
provider.watchAuditRequests((result) => {
  console.log(result.passed ? "PASSED" : "FAILED");
});

Phase 4: Fraud Detection

Users verify receipts locally:

const result = await verifyReceipt(receipt);

if (result.status === "FRAUD_NON_INCLUSION") {
  // Provider never batched the query
  disputeNonInclusion(receipt);
}

if (result.status === "FRAUD_INVALID_PROOF") {
  // Provider tampered with batch data
  disputeFraudulentInclusion(batchId, seqNum);
}

---

Economic Security

Parameter	Value
Provider Stake	0.0001 ETH
Audit Stake	0.00005 ETH
Dispute Stake	0.0001 ETH
Audit Deadline	24 hours
Grace Period	1 hour

Slashing conditions:

• Failed audit → stake to challenger
• Missed deadline → slashed
• Valid dispute → slashed, user compensated