Disrupting Facial Recognition

#!/usr/bin/env python3
# Prerequisites: pip install deepface opencv-python tensorflow
"""Compare two face images and measure verification confidence.
Use this to test how appearance changes affect match scores."""
# DeepFace returns cosine distance by default (0 = identical, 1 = completely different)
from deepface import DeepFace
import json

# Test baseline (same person, normal conditions)
baseline = DeepFace.verify(
    img1_path="baseline_front.jpg",
    img2_path="baseline_angle.jpg",
    model_name="ArcFace",
    detector_backend="retinaface"
)
print("Baseline match:", json.dumps(baseline, indent=2))

# Test mitigated (same person, with countermeasures)
mitigated = DeepFace.verify(
    img1_path="baseline_front.jpg",
    img2_path="mitigated_hat_glasses.jpg",
    model_name="ArcFace",
    detector_backend="retinaface"
)
print("Mitigated match:", json.dumps(mitigated, indent=2))
# Positive delta = mitigation INCREASED distance from reference (good — harder to match)
print(f"Confidence delta: {mitigated['distance'] - baseline['distance']:.4f}")

# Expected output:
# === Baseline Verification ===
# Model: VGG-Face | Distance: 0.2134 | Threshold: 0.40 | Verified: True
# Model: Facenet  | Distance: 0.7891 | Threshold: 1.10 | Verified: True
#
# === Mitigated Verification ===
# Model: VGG-Face | Distance: 0.5847 | Threshold: 0.40 | Verified: False  ← defense effective
# Model: Facenet  | Distance: 1.3204 | Threshold: 1.10 | Verified: False  ← defense effective
#
# Confidence delta (VGG-Face): +0.3713 (higher = better mitigation)

# Install Fawkes image cloaking tool
# ⚠ The pip package is broken since late 2023 — install from source:
git clone https://github.com/Shawn-Shan/fawkes.git && cd fawkes
pip install -e .

# Cloak images before uploading to social media
# This adds imperceptible perturbations that disrupt model training
# -m modes: low (fastest, minimal perturbation), mid (balanced), high (strongest cloak, slower)
fawkes -d ./photos_to_upload/ -m high --gpu 0

# Verify cloaking was applied
fawkes -d ./photos_to_upload/ --mode verify

#!/usr/bin/env python3
# Prerequisites: pip install mediapipe opencv-python
"""Test face detection rates under various conditions.
Measures how countermeasures affect detector reliability."""
import cv2
import mediapipe as mp
import os
import json

mp_face = mp.solutions.face_detection
results = {}

conditions = {
    "baseline": "captures/baseline/",
    "hat_only": "captures/hat/",
    "glasses_only": "captures/glasses/",
    "hat_and_glasses": "captures/hat_glasses/",
    "hat_glasses_angle": "captures/hat_glasses_angle/",
    "ir_glasses": "captures/ir_glasses/",
    "patterned_scarf": "captures/patterned_scarf/"
}

with mp_face.FaceDetection(
    model_selection=1,           # model_selection: 0 = short-range (faces within 2m), 1 = full-range (up to 5m)
    min_detection_confidence=0.5 # min_detection_confidence: 0.5 = balanced threshold (lower catches more faces but more false positives)
) as detector:
    for condition, path in conditions.items():
        detected = 0
        total = 0
        confidences = []
        for img_file in os.listdir(path):
            img = cv2.imread(os.path.join(path, img_file))
            if img is None:
                continue
            rgb = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
            result = detector.process(rgb)
            total += 1
            if result.detections:
                detected += 1
                confidences.append(result.detections[0].score[0])
        
        avg_conf = sum(confidences) / len(confidences) if confidences else 0
        results[condition] = {
            "detection_rate": f"{detected/total*100:.1f}%" if total > 0 else "N/A",
            "avg_confidence": f"{avg_conf:.3f}",
            "samples": total
        }

print(json.dumps(results, indent=2))

# --- Expected Output ---
# {
#   "baseline": {
#     "detection_rate": "100.0%",
#     "avg_confidence": "0.943",
#     "samples": 25
#   },
#   "hat_only": {
#     "detection_rate": "88.0%",
#     "avg_confidence": "0.812",
#     "samples": 25
#   },
#   "glasses_only": {
#     "detection_rate": "96.0%",
#     "avg_confidence": "0.887",
#     "samples": 25
#   },
#   "hat_and_glasses": {
#     "detection_rate": "72.0%",
#     "avg_confidence": "0.694",
#     "samples": 25
#   },
#   "ir_glasses": {
#     "detection_rate": "32.0%",
#     "avg_confidence": "0.531",
#     "samples": 25
#   },
#   "patterned_scarf": {
#     "detection_rate": "44.0%",
#     "avg_confidence": "0.578",
#     "samples": 25
#   }
# }

#!/usr/bin/env python3
# Prerequisites: pip install deepface numpy
"""Analyze face embedding stability across conditions.
Lower cosine similarity = harder to match across captures."""
from deepface import DeepFace
import numpy as np

def get_embedding(img_path, model="ArcFace"):
    result = DeepFace.represent(img_path, model_name=model, detector_backend="retinaface")
    return np.array(result[0]["embedding"])

def cosine_similarity(a, b):
    return np.dot(a, b) / (np.linalg.norm(a) * np.linalg.norm(b))

# Reference embedding (enrollment photo)
ref = get_embedding("reference_portrait.jpg")

# Test conditions
test_images = {
    "normal_lighting": "test_normal.jpg",
    "low_light": "test_lowlight.jpg",
    "hat_brim": "test_hat.jpg",
    "reflective_glasses": "test_reflective.jpg",
    "hat_plus_glasses": "test_hat_glasses.jpg",
    "angle_30_deg": "test_angle30.jpg",
    "distance_10m": "test_distance10m.jpg",
}

print(f"{'Condition':<25} {'Cosine Sim':>12} {'Would Match':>12}")
print("-" * 52)
for condition, path in test_images.items():
    try:
        emb = get_embedding(path)
        sim = cosine_similarity(ref, emb)
        # 0.68 = commonly used cosine similarity threshold for face match (model-dependent; VGG-Face typical range: 0.60–0.75)
        match = "YES" if sim > 0.68 else "NO"
        print(f"{condition:<25} {sim:>12.4f} {match:>12}")
    except Exception as e:
        print(f"{condition:<25} {'FAILED':>12} {'N/A':>12}")

# Expected output:
# Condition          | Similarity | Match
# ────────────────────────────────────────
# baseline           |     0.9234 | YES
# hat_sunglasses     |     0.7102 | YES
# ir_leds_active     |     0.5843 | NO   ← effective
# cv_dazzle_makeup   |     0.4127 | NO   ← effective
# fawkes_cloaked     |     0.6312 | NO   ← effective