aitbc/docs/development/zk-circuits.md

ZK Circuits Engine

Overview

The ZK Circuits Engine provides zero-knowledge proof capabilities for privacy-preserving machine learning operations on the AITBC platform. It enables cryptographic verification of ML computations without revealing the underlying data or model parameters.

Architecture

Circuit Library

• ml_inference_verification.circom: Verifies neural network inference correctness
• ml_training_verification.circom: Verifies gradient descent training without revealing data
• receipt_simple.circom: Basic receipt verification (existing)

Proof System

• Groth16: Primary proving system for efficiency
• Trusted Setup: Powers-of-tau ceremony for circuit-specific keys
• Verification Keys: Pre-computed for each circuit

Circuit Details

ML Inference Verification

pragma circom 2.0.0;

template MLInferenceVerification(INPUT_SIZE, HIDDEN_SIZE, OUTPUT_SIZE) {
    signal public input model_id;
    signal public input inference_id;
    signal public input expected_output[OUTPUT_SIZE];
    signal public input output_hash;

    signal private input inputs[INPUT_SIZE];
    signal private input weights1[HIDDEN_SIZE][INPUT_SIZE];
    signal private input biases1[HIDDEN_SIZE];
    signal private input weights2[OUTPUT_SIZE][HIDDEN_SIZE];
    signal private input biases2[OUTPUT_SIZE];

    signal private input inputs_hash;
    signal private input weights1_hash;
    signal private input biases1_hash;
    signal private input weights2_hash;
    signal private input biases2_hash;

    signal output verification_result;
    // ... neural network computation and verification
}

Features:

• Matrix multiplication verification
• ReLU activation function verification
• Hash-based privacy preservation
• Output correctness verification

ML Training Verification

template GradientDescentStep(PARAM_COUNT) {
    signal input parameters[PARAM_COUNT];
    signal input gradients[PARAM_COUNT];
    signal input learning_rate;
    signal input parameters_hash;
    signal input gradients_hash;

    signal output new_parameters[PARAM_COUNT];
    signal output new_parameters_hash;
    // ... gradient descent computation
}

Features:

• Gradient descent verification
• Parameter update correctness
• Training data privacy preservation
• Convergence verification

API Integration

Proof Generation

POST /v1/ml-zk/prove/inference
{
  "inputs": {
    "model_id": "model_123",
    "inference_id": "inference_456",
    "expected_output": [2.5]
  },
  "private_inputs": {
    "inputs": [1, 2, 3, 4],
    "weights1": [0.1, 0.2, 0.3, 0.4],
    "biases1": [0.1, 0.2]
  }
}

Proof Verification

POST /v1/ml-zk/verify/inference
{
  "proof": "...",
  "public_signals": [...],
  "verification_key": "..."
}

Development Workflow

Circuit Development

1. Write Circom circuit with templates
2. Compile with circom circuit.circom --r1cs --wasm --sym --c -o build/
3. Generate trusted setup with snarkjs
4. Export verification key
5. Integrate with ZKProofService

Testing

• Unit tests for circuit compilation
• Integration tests for proof generation/verification
• Performance benchmarks for proof time
• Memory usage analysis

Performance Characteristics

• Circuit Compilation: ~30-60 seconds
• Proof Generation: <2 seconds
• Proof Verification: <100ms
• Circuit Size: ~10-50KB compiled
• Security Level: 128-bit equivalent

Security Considerations

• Trusted Setup: Powers-of-tau ceremony properly executed
• Circuit Correctness: Thorough mathematical verification
• Input Validation: Proper bounds checking on all signals
• Side Channel Protection: Constant-time operations where possible

Future Enhancements

• PLONK/STARK Integration: Alternative proving systems
• Recursive Proofs: Proof composition for complex workflows
• Hardware Acceleration: GPU-accelerated proof generation
• Multi-party Computation: Distributed proof generation