The Logician.Facia.AI · Case study
Tested liveness systems with physical masks, deepfakes and injection attacks. The findings informed stronger detection and new product directions.
When I joined as a Product Analyst, Facia already had a mature biometric security product with liveness detection defending against 53+ spoofing attack types. My mandate was distinct: I was not asked to build features. I was asked to break the product. Think like a fraudster, find every way the system could fail, and document vulnerabilities before clients or real attackers discovered them.
This is the defining story. During adversarial testing, I escalated through three phases of attack complexity: physical presentation attacks (printed photos, screen replays, 3D masks), generative AI attacks (deepfakes, face swaps, face morphs using 40+ tools), and injection attacks (virtual camera injection, API-level synthetic image submission).
One of the earliest and most consequential discoveries was also the simplest. A $2 nylon stocking mask successfully spoofed the liveness detection. The mask preserved enough facial geometry to pass depth analysis while defeating texture detection. This wasn't just a Facia vulnerability. I tested the same attack against Amazon Rekognition, one of the most widely deployed facial recognition services in the world. It worked. I also spoofed BioID's liveness detection.
I documented the AWS finding publicly on LinkedIn. The point wasn't to embarrass a vendor. It was to demonstrate an industry-wide gap. If a $2 mask can defeat a billion-dollar cloud provider's liveness check, the entire industry needs to rethink its approach to presentation attack detection.
Thousands of deepfake images, face swaps, face morphs, AI-generated faces, and manipulated videos using 40+ generative AI tools. Each successful spoof was documented with exact conditions. This dataset became the foundation for two new product lines: Deepfake Detection (evaluated using Meta's Deepfake Detection Challenge Dataset of 124,000 videos) and AI Image Detection (detects AI-generated images via colour inconsistencies, lighting anomalies, and metadata).
Rigorous testing identified systematic inaccuracies across demographics. Led in-house consented dataset creation with diverse age ranges and ethnicities. The result was an 80% improvement in estimation accuracy for Challenge 21 and Challenge 25 compliance.
$2 mask spoofed AWS Rekognition, publicly documented
Adversarial dataset enabled Deepfake Detection and AI Image Detection product lines
80% improvement in age estimation accuracy
60% reduction in manual fraud reviews
Python automation cut verification testing time by 50%
This period crystallised my approach. Most PMs focus on building. I learned to break first. The adversarial mindset, systematically escalating attacks, documenting failure conditions, and strengthening the product, became my lens for every product decision. When I later built AML Watcher's risk scoring or Barie's hallucination minimisation, I was applying the same principle: find the failure before the user does.