Discoveries How It Works Metrics Full Stats
Research Network Active (0 online)

Could a pattern on your clothing fool Facial RecognitionFacial RecognitionFacial Recognition?

This is a distributed research project discovering adversarial patterns that fool AI surveillance systems. Privacy shouldn't be a luxury. It's a fundamental right.

0 Tests
0 Anomalies
0 Multi-Model
0 Extreme
Live Discovery View All →
Found by Loading...

Why This Research Matters

Do you know how many times a day you're subjected to facial recognition? The camera at your ATM. The gas pump. Every doorbell in your neighborhood. Airports, shopping centers, and city streets. You are being tracked, identified, and cataloged by systems you never consented to. We believe there's another way. Privacy is not a luxury. It is a fundamental right.

Building on Giants

Pioneering work by Adam Harvey, Capable, and Adversarial Fashion proved that clothing could disrupt surveillance. Their groundbreaking research inspired this project. We're continuing that evolution with modern models and rigorous methodology.

Real World Models

This is not academic research against outdated benchmarks. We test against the exact models deployed by Clearview AI, law enforcement, and commercial surveillance systems. Older adversarial patterns were fragile against modern AI. Our 10-model gauntlet ensures patterns work against today's systems.

Open Research

Testing billions of pattern combinations requires massive computational power. Our distributed network lets anyone contribute spare GPU cycles to this research. Access is restricted to non-commercial research, academic, and public interest use. Those affiliated with surveillance companies like Clearview AI, Palantir, or Voyager Labs are explicitly excluded.

From Pattern to Discovery

Our distributed network tests thousands of visual patterns against state-of-the-art facial recognition models.

[1] Generate Patterns

61+ pattern types exploiting AI vision weaknesses through geometric, frequency, and adversarial techniques.

[2] Scientific Testing

Two-stage methodology: establish baseline, then measure pattern impact on detection accuracy.

[3] Model Gauntlet

Distributed workers test against 10 models: 4 person detectors, 4 face detectors, 2 recognizers.

[4] Evolve & Record

Successful patterns breed new generations through crossover and mutation algorithms.

Each pattern "recipe" exploits specific weaknesses in AI vision systems. These are not random noise — they're a library of 61+ specific attack techniques:

Geometric Optical illusions and dazzle camouflage creating visual interference
Glitch Art Pixel sorting and frequency-domain FFT noise disruption
Adversarial ML PGD & DCA gradient attacks optimized against model weights
saliency_eye_attack Overloads the system with fake eye features
dazzle_surgical_lines Breaks up facial geometry between key landmarks
📚 Patterns are layered and combined to create unique test candidates. View the complete pattern library →

Every test follows a rigorous two-stage protocol ensuring reproducible, scientifically valid results:

Establish Baseline

Original "Persona" image processed through all 10 models.

Person detection counts Face detection counts Recognition confidence scores Identity matches

Test Pattern Effect

Pattern applied via green-screen compositing. Identical model suite re-runs tests.

Success = consistent, measurable degradation Changes must be reproducible, not random

Standardized test images featuring high-quality, diverse synthetic faces — different descents, genders, and ages — ensuring testing covers vulnerable populations who need protection most.

Clients operate as a coordinated network, contributing compute power simultaneously. Results are aggregated in real-time, dramatically accelerating research.

Work Distribution Batches assigned based on worker compute capabilities
Parallel Processing Workers process patterns locally on their own hardware
Live Aggregation Discoveries submitted and tracked on the live dashboard

Distributed workers test patterns against 10 models representing real-world commercial surveillance:

P1 YOLOv8n
Axon Body 4/Fleet 3 body cameras
Next-gen law enforcement; requires "higher accuracy in low light"
P2 YOLOv5s
Axis P1465-LE, Q1656-LE cameras
Explicitly benchmarked in Axis's public GitHub repo
P3 SSD-MobileNetV2
Hikvision AcuSense & Axis Pro Gen 2
Lightweight standard in millions of retail/office cameras
P4 ResNet34-SSD
Avigilon & Palantir AI NVR
NVIDIA PeopleNet — detects people, faces, bags from long distances
F1 InsightFace Buffalo_L
Dahua WizSense & Hikvision MinMoe
High-accuracy for office access; prevents spoofing attacks
F2 FaceNet
Axon Redaction Assistant
Auto face blurring via NXP FaceNet512 optimization
F3 MTCNN
Legacy Smart City & Intel OpenVINO
Classic detector still in traffic monitoring systems
F4 RetinaFace
Stadiums & airports
State-of-art for detecting small faces in large crowds
R1 ArcFace
Clearview AI & state surveillance
Current open-source SOTA; foundation for aggressive ID platforms
R2 FaceNet
FBI & law enforcement databases
Legacy government systems built 5-10 years ago

Successful patterns don't just get recorded — they breed new generations through genetic evolution:

Genetic Algorithm Winners flagged to PRIORITY_TESTS. Crossover combines best parts of successful patterns. Mutation adds small random tweaks. Evolved recipes deploy back to workers.
Epoch Cycle One complete iterative cycle: genetic algorithm generates evolved patterns → workers test them → results feed next epoch. Epoch count tracks generational progress.
Multi-Model Breakthroughs
+1000 TOTAL_STEALTH All P+F models defeated +400 PERSON_STEALTH All P1-P4 defeated +400 FACE_STEALTH All F1-F4 defeated +200 EXTREME Any P + any F defeated
PRIORITY triggered by 2+ model anomalies
Person Detection
+200 PERSON_LOST Fewer than baseline +100 EXTRA_PERSONS Spurious detections
Face Detection
+200 NO_FACES Zero faces found +200 PRIMARY_LOST Main face gone +100 EXTRA_FACES Spurious faces +100 LOW_CONF Low confidence match
Recognition
+100 RECOG_FAIL Wrong identity

When patterns bypass 2+ models: auto-generate 300 DPI pattern images, save test images, submit recipe data for administrator review.

Network Performance

Real-time statistics from our distributed research network.

0
Total Tests
0
Anomalies Found
0
Multi-Model Defeats
0
Extreme Anomalies
0
Current Epoch

Top Performing Patterns

Most Effective Combinations

Top Contributors

Recognizing the researchers powering our distributed network.

Rank Researcher Tests Success Rate
Loading leaderboard...

Ready to Contribute?

Join thousands of researchers worldwide. Your GPU can help discover patterns that protect privacy and improve AI safety.

Get Started Now

Welcome Back

Sign in to track your contributions and access your profile