The engineering document behind the MyShape Protocol. Motion Vector format, PES engine, proof system, SST topology, and reference implementation.
Camera captures motion at 30fps. MediaPipe Pose extracts 33 landmarks.
SST mapper converts to standardized 18-point topology (§2.2).
Feature Pipeline (§3): Preprocessing → Temporal → Kinematic → Entropy → Hash.
Output: Motion Vector — a fixed-size geometric signature, not raw video.
The raw signal never leaves the device. Only the hashed vector is transmitted.
4-dimensional Presence Entropy Score (PES §3.5):
μTiming Variance — micro-second frame timing irregularities (biological).
Noise Residual — signal-to-noise ratio from low-pass prediction error.
Frequency Entropy — Shannon entropy over DFT power spectrum.
Biological Perturbation — cross-joint jerk correlation (6 joint pairs).
Combined via weighted sum: PES = 0.20·μT + 0.25·ε + 0.20·Hf + 0.35·Bio.
False Accept Rate (FAR): empirically < 5% for AI-generated motion (PES threshold 0.65).
False Reject Rate (FRR): < 15% for real humans in normal standing conditions.
Entropy Gap Theorem (§10) proves: Pr[AI PES ≥ 0.65] → 0.
Note: These are protocol-level metrics, not conventional FAR/FRR. MyShape verifies presence, not identity.
Minimum: 640×480 camera, 20fps, any modern browser (Firefox recommended).
Recommended: 1280×720, 30fps, dedicated webcam.
Processing: all on-device. No GPU required. No cloud dependency.
Latency: < 100ms for feature extraction. < 10ms for proof verification.
Proof size: < 512 bytes (ZK-Presence composite).
Platform support: Web (TypeScript SDK). iOS/Android/Unity/Unreal (planned).
8 attack signatures detected across 4 PES dimensions (§5):
Generative AI — frequency entropy collapse + missing jerk correlation.
Replay — uniform inter-frame timing + stale hash.
Imitation — cannot replicate unconscious micro-motion patterns.
Mocap — over-clean data lacks biological noise.
Sensor spoofing — inter-frame consistency checks + frequency monitoring.
Adversarial pose injection — MV normalization removes perturbations.
Statistical forgery — cannot satisfy all 4 dimensions simultaneously.
Digital twin — lacks original neural noise source.
Defense-in-depth: 5-layer architecture with independent security boundaries.
16 protocol engines implemented in TypeScript. Zero external dependencies. All processing on-device. No data stored or transmitted.
View SDK →