ZK-Sentinel V11
TESTNET
🧠 ZKML Integration
Halo2-based machine learning for coercion detection
Architecture
🎙️
Voice / Text
Input capture
📊
Feature Extract
MFCC / embeddings
🧠
LSTM + MLP
Stress classifier
🔒
EZKL → Halo2
ZK proof of inference
⛓️
On-chain Verify
Halo2Verifier.sol
Model Specifications
| Parameter | Value |
|---|---|
| Architecture | LSTM (32 units) → Dense(16) → Dense(1, sigmoid) |
| Input | 13 MFCC features × 20 time steps |
| Output | Stress probability [0.0 — 1.0] |
| Threshold | 0.7 (configurable per pool) |
| Training Data | RAVDESS + CREMA-D + custom corpus |
| Quantization | INT8 via EZKL for Halo2 compatibility |
| Proof Size | ~1.2 KB (Halo2 KZG commitment) |
| Verification Gas | ~280K gas on-chain |
Privacy Guarantee: The ZKML proof demonstrates that the ML model classified the user as "not under stress" — without revealing the raw biometric input. The verifier only sees the boolean result, never the voice/text data.