Privacy Risks of Facial Recognition Technology

Facial recognition has moved from border checkpoints to retail stores, concert venues, apartment buildings, and hiring platforms — often without meaningful consent. Unlike a password you can change, your face is permanently attached to your identity. This guide covers where FRT is deployed, what the risks are, your legal options, and technical countermeasures.

How Facial Recognition Works

Modern FRT uses deep neural networks to generate a 128–512 dimensional embedding vector from a face image. Two faces match if their embedding vectors are within a threshold distance. The quality of the embedding depends on image resolution, lighting, and model training data.

Input image → Face detection → Face alignment → Feature extraction → Embedding vector
                                                                           ↓
                                                              Compare with database
                                                                           ↓
                                                              Match / No Match (+ confidence)

Key specifications that affect privacy risk:

• False Acceptance Rate (FAR): probability of matching the wrong person — matters for wrongful identification
• False Rejection Rate (FRR): probability of failing to match the right person
• Training data demographics: systems trained on non-representative data have higher FAR for certain ethnic groups (documented by NIST FRVT studies)

Where You Are Being Scanned

1. Public Infrastructure

Law enforcement use of real-time FRT on CCTV networks is expanding:

• UK: the Metropolitan Police uses NEC NeoFace live on CCTV feeds in London
• China: Skynet/Sharp Eyes system covers hundreds of millions of cameras
• US: FBI’s NGI-IPS (Next Generation Identification) matches against 650M+ face images
• EU: the AI Act (2024) prohibits real-time biometric surveillance in public spaces with exceptions for law enforcement

2. Retail and Event Venues

Surveillance companies like Clearview AI, Facewatch, and FaceFirst sell FRT to retailers to identify known shoplifters — and in some cases, anyone on a “watchlist” maintained by the vendor. You may be scanned when entering a supermarket, stadium, or hotel.

3. Hiring and Remote Proctoring

Platforms like HireVue, Proctorio, and ExamSoft analyze facial expressions and “micro-expressions” during job interviews and exams. Vendors claim these predict candidate quality; independent research has found no validity for these claims.

4. Phone Unlock and Account Recovery

Face ID (Apple), Face Unlock (Android), and similar systems use 3D depth sensors with liveness detection. The biometric is processed on-device (Secure Enclave / TEE) and never leaves the device in raw form. Risk is comparatively low but law enforcement can compel biometric unlock in many jurisdictions, unlike PINs.

5. Social Media Photo Tagging

Every time you’re tagged in a photo or use a filter, the platform builds a facial embedding linked to your profile. Facebook’s DeepFace system accumulated embeddings for 1.4 billion faces before being ordered to delete the data.

The Clearview AI Problem

Clearview AI scraped over 50 billion photos from social media, news, and public websites to build the largest commercial FRT database. Law enforcement customers can upload a face photo and get matches across the internet.

Legal status in 2026:

• EU: ruled illegal under GDPR; Clearview banned from processing EU residents’ data (€20M fine, 2023)
• Australia: ruled a breach of the Privacy Act; ordered to delete Australian data
• US: no federal prohibition; some state-level restrictions (Illinois BIPA)
• UK: ICO issued enforcement notice; Clearview faces ongoing litigation

Technical Countermeasures

Adversarial Makeup and Clothing

Research shows that specific infrared-reflective patterns and adversarial makeup configurations can cause FRT misidentification. These are partially effective in controlled conditions:

IR-reflective patches:
- Embedded in eyeglass frames (e.g., Reflectacles brand)
- Block infrared illumination used by many FRT systems at night

CV Dazzle makeup (Adam Harvey):
- Asymmetric patterns that break the facial symmetry detectors use
- Less effective against deep learning models than against earlier Haar cascade detectors

Adversarial t-shirts and hoodies:
- Printed patterns that trigger object classification interference
- Effective against some person-detection networks (YOLO variants)

Effectiveness is limited: video analysis can extract clean frames between distorted ones, and adversarial patterns effective against one model may fail against another.

Technical Opt-Out from Photo Platforms

# Request deletion of face templates from major platforms

# Facebook/Meta — biometric data deletion (where applicable under BIPA)
# Settings > Privacy > Face Recognition > Manage Face Recognition Data

# Google Photos
# Settings > Group Similar Faces > Off (prevents further processing)
# Takeout + account deletion to remove existing embeddings

# LinkedIn
# Settings > Data Privacy > Profile Data for Targeted Advertising

Submit GDPR/CCPA Erasure Requests

In the EU and California, you have the right to request deletion of biometric data:

# Template for GDPR Article 17 erasure request (EU)
EMAIL_TEMPLATE = """
Subject: GDPR Article 17 - Right to Erasure - Biometric Data

To whom it may concern,

Under Article 17 of the General Data Protection Regulation (GDPR), I request
the immediate erasure of all biometric data you hold about me, including but
not limited to facial recognition embeddings, voiceprints, and any derived
inferences.

My identifying information:
Name: [YOUR NAME]
Email used on your platform: [YOUR EMAIL]
Any user IDs or profile URLs: [IDs]

I request:
1. Confirmation that all biometric data has been deleted
2. Confirmation that data has not been shared with third parties
3. If shared, the identity of all recipients

Please respond within 30 days as required by Article 12 GDPR.

Regards,
[YOUR NAME]
"""

Legal Framework by Jurisdiction

What Liveness Detection Can and Cannot Stop

Modern FRT systems use liveness detection to prevent spoofing with photos:

• 2D liveness: detects flat photos (blinking, head movement challenges)
• 3D liveness: uses depth sensors; defeated by high-quality 3D-printed masks
• IR liveness: detects IR reflection patterns from living tissue

None of these prevent identification of a real person’s face. They only prevent spoofing (presenting a fake face to authenticate as someone else).

Reducing Your Biometric Footprint

1. Reduce photos indexed publicly: audit and remove tag yourself from photos on social platforms
2. Request opt-out from data brokers: services like Spokeo, BeenVerified, and PimEyes index public photos for face search
3. Use privacy-respecting video calls: avoid platforms that claim to analyze facial expressions during calls
4. Disable phone face unlock in high-risk situations: use PIN when crossing borders or attending protests
5. Use a VPN when accessing platforms: reduces IP-geolocation correlation with your face data

Related Reading

• Android Privacy Best Practices 2026
• Privacy-Focused Instant Messaging Comparison 2026
• How to Prevent Someone from Tracking Your Location

Built by theluckystrike — More at zovo.one