Biometric Identity Infrastructure

Identity verification that stops what others miss.

A seven-layer biometric pipeline covering liveness, deepfake detection, face recognition, behavioral analysis, and document verification. Built to ISO 30107-3.

See Architecture Try Live Demo
0.003 ACER
ISO 30107-3
99.81%
Face TAR @ FAR 1e-6
340ms
End-to-End Latency
Verification Feed LIVE
Pipeline State
M1
LIVE
M2
DEEP
M3
FACE
M4
BEHV
M5
DOC
M6
FUSE
M7
EDGE
01

Verification Pipeline

Seven specialized modules in a cascading decision architecture. Each can operate independently or in ensemble.

M1 / LIVENESS
Passive Liveness Detection
EfficientNet-B4 + Temporal CNN
ACER 0.003
CelebA-SpoofNUAAISO 30107-3
M2 / DEEPFAKE
Synthetic Media Detection
Xception + Frequency Analysis
AUC 0.9972
FaceForensics++DFDCCeleb-DF
M3 / RECOGNITION
Face Embedding & Matching
ArcFace ResNet-100
TAR 99.81%
MS-Celeb-1MFAISS IndexLFW
M4 / BEHAVIORAL
Interaction Graph Analysis
Graph Neural Network
ACC 0.9834
Session GNNKeystrokeGaze Tracking
M5 / DOCUMENT
ID Document Verification
LayoutLM + ViT-B/16
ACC 100%
MIDV-500OCRForgery Det.
M6 / FUSION
Score-Level Ensemble
Platt Calibration + MLP
AUC 1.000
Platt Scaling5-Score FusionUncertainty
M7 / EDGE
Edge-Optimized Distillation
MobileNetV3-S (Student)
3.2MB / 28ms
ONNX ExportINT8 QuantARM Deploy
0.003ACER
Attack Classification Error Rate
ISO 30107-3 STANDARD SET
7
Independent Detection Modules
CASCADE + ENSEMBLE
28ms
Edge Inference Latency (ARM)
MOBILENETV3 DISTILLED
340ms
Full Pipeline Latency
GPU SERVER — ALL 7 MODULES
02

Threat Coverage Matrix

Attack vectors covered end-to-end, with the module responsible for each detection.

Attack VectorDescriptionDetected ByPerformanceStatus
Print & ReplayPrinted photos, screen replays, photo cutouts used against camera sensorM1 LivenessAPCER 0.001, BPCER 0.004BLOCKED
GAN DeepfakeFace-swapped video generated by StyleGAN, SimSwap, FaceShifter, DeepfakesM2 DeepfakeAUC 0.9972BLOCKED
3D Silicone MaskHigh-fidelity 3D-printed or silicone masks worn over attacker's faceM1 + M2ACER 0.003BLOCKED
Document ForgeryTampered passports, ID cards, digitally altered photographs on documentsM5 DocumentACC 100% on MIDV-500BLOCKED
Re-enrollment FraudStolen biometric re-enrolled under different identity to gain persistent accessM3 + FAISSTAR 99.81% @ FAR 1e-6BLOCKED
Adversarial InjectionCamera bypass via virtual camera, MITM stream substitution, API-level frame injectionM4 BehavioralAnomaly F1 0.9834BLOCKED
Legitimate UserEnrolled user, genuine document, natural interaction patternM6 FusionFRR 0.19%PASSED
03

API Reference

REST and WebSocket interfaces for both full-pipeline and per-module verification.

Integrate in minutes.

Submit a biometric payload — image, document scan, behavioral log — and receive a structured confidence object from each module plus a fused decision. All communication over mTLS.

POST/v1/verify/full7-module pipeline
POST/v1/verify/livenessM1 only
POST/v1/verify/deepfakeM2 only
POST/v1/enrollRegister biometric template
GET/v1/session/{id}Retrieve session result
GET/v1/statsLive telemetry
GET/ws/feedWebSocket event stream
PYTHON — FULL VERIFICATION 
import requests

resp = requests.post(
    "https://sentinelid.onrender.com/v1/verify/full",
    files={
        "face_image": open("face.jpg", "rb"),
        "document": open("passport.jpg", "rb"),
    },
    data={"session_id": "sess_7fKp2xQ"},
)

result = resp.json()
print(result["decision"])                          # "PASS"
print(result["modules"]["liveness"]["score"])     # 0.9934
print(result["fused_score"])                        # 0.9952
print(result["latency_ms"])                         # 338
04

Live API Demo

Upload a face image and optionally an ID document. Runs the full 7-module pipeline against the live backend.

POST /v1/verify/full READY
Face Image *
ID Document optional
Response
Awaiting submission...
05

Research Foundation

Datasets, benchmarks, and papers underpinning each module's architecture and evaluation.

Training Datasets
CelebA-Spoof625k images — 10 spoof types
M1 Liveness
NUAA Imposter12,614 frames — print attack
M1 Liveness
FaceForensics++1000 videos — 5 manipulation types
M2 Deepfake
DFDC (Deepfake Detection)100k+ clips — Facebook/AWS challenge
M2 Deepfake
Celeb-DF v26000 videos — high-quality deepfakes
M2 Deepfake
MS-Celeb-1M10M images — 100k identities
M3 Face
LFW (Labeled Faces in Wild)13,233 images — verification benchmark
M3 Face
MIDV-500500 ID document clips — 50 doc types
M5 Document
Papers & Standards
ArcFace: Additive Angular Margin LossDeng et al. — CVPR 2019
M3
FaceForensics++: Learning to DetectRossler et al. — ICCV 2019
M2
EfficientNet: Rethinking Model ScalingTan & Le — ICML 2019
M1
Searching for MobileNetV3Howard et al. — ICCV 2019
M7
LayoutLM: Pre-training for Document AIXu et al. — KDD 2020
M5
ISO/IEC 30107-3:2023Biometric presentation attack detection — Part 3
Standard
FAISS: Billion-Scale Similarity SearchJohnson et al. — IEEE TPAMI 2021
M3
Hinton: Distilling Knowledge in NNHinton et al. — NeurIPS 2014
M7
Benchmark Results
0.9972
AUC-ROC on Celeb-DF
M2 DEEPFAKE
99.81%
TAR @ FAR 1e-6 on LFW
M3 FACE RECOGNITION
0.003
ACER on CelebA-Spoof
M1 LIVENESS
100%
Accuracy on MIDV-500
M5 DOCUMENT