Face Verification System (1:1) Development

Development of Face Verification Systems (1:1)

Face verification answers the question "is this the same person?". Unlike identification (1:N database search), verification compares two specific images and returns True/False. Main applications: identity verification during online registration, matching document photo with selfie, mobile app authentication.

Verification Algorithm

import numpy as np
from insightface.app import FaceAnalysis

class FaceVerifier:
    def __init__(self, threshold: float = 0.5):
        self.app = FaceAnalysis(
            providers=['CUDAExecutionProvider', 'CPUExecutionProvider']
        )
        self.app.prepare(ctx_id=0, det_size=(640, 640))
        self.threshold = threshold  # adjusted to FAR/FRR requirements

    def verify(self, image1: np.ndarray,
               image2: np.ndarray) -> dict:
        face1 = self._extract_face(image1)
        face2 = self._extract_face(image2)

        if face1 is None or face2 is None:
            return {'verified': False, 'reason': 'face_not_detected'}

        # Cosine similarity between ArcFace embeddings
        similarity = self._cosine_similarity(face1.embedding, face2.embedding)

        return {
            'verified': similarity >= self.threshold,
            'similarity': float(similarity),
            'threshold': self.threshold
        }

    def _cosine_similarity(self, a: np.ndarray, b: np.ndarray) -> float:
        return float(np.dot(a, b) / (np.linalg.norm(a) * np.linalg.norm(b)))

Threshold Tuning: FAR vs FRR

FAR (False Accept Rate) — probability of accepting stranger as legitimate. Critical for security. FRR (False Reject Rate) — probability of rejecting legitimate user. Affects UX.

These metrics are opposing: lowering FAR raises FRR. Threshold choice depends on application:

EER (Equal Error Rate) — point where FAR = FRR. For ArcFace on LFW: EER ≈ 0.17%.

Liveness Detection

Without anti-spoofing, verification is vulnerable to attacks: photo on screen, 3D mask. Required components:

Passive liveness check — analyze skin texture (LBP, FrequentNet), detect screen artifacts:

from silent_face_anti_spoofing import AntiSpoof

anti_spoof = AntiSpoof(model_path='2.7_80x80_MiniFASNetV2.pth')

def check_liveness(face_crop: np.ndarray) -> dict:
    prediction = anti_spoof.predict(face_crop)
    return {
        'is_real': prediction['label'] == 1,
        'score': prediction['probability']
    }

Active liveness check — user performs random action: blink, head turn, speak digit. Sequence of frames is validated.

Document Verification (KYC)

For Know Your Customer tasks: compare document photo (passport, driver's license) with selfie.

Specifics: document photos are often low quality, may be scans with watermarks, different shooting conditions. Document photo preprocessing: detect photo zone, fix perspective, normalize brightness.

ArcFace verification accuracy matching document↔selfie: 94–97% TAR@FAR=0.1%.

Metrics and Benchmarks

• LFW (Labeled Faces in the Wild): academic standard. ArcFace: 99.83%
• IJB-B/IJB-C: harder datasets with video. ArcFace TAR@FAR=1e-4: 94.0/96.5%
• MegaFace Challenge: 1M distractors. ArcFace Rank-1: 98.35%