How to Detect When the Probe in 1-to-N Facial ID is Out-of-Gallery.
The short version: A biometrics researcher's new talk tackles a core flaw in police facial ID: algorithms that always return a 'match' even when the real person isn't in the database.
Our Take
Credit where it's due: ND professor Kevin W. Bowyer is naming a problem that police departments and vendors have spent years glossing over. One-to-many facial identification doesn't just find matches — it's built to always return one, ranked by confidence, whether or not the actual person is anywhere in the gallery. Bowyer's talk walks through documented wrongful arrests tied to this workflow and previews research aimed at teaching systems to recognize when the true probe identity simply isn't in the database at all.
That's a real technical contribution, but it's worth sitting with what it implies: after years of headline wrongful arrests, the fix being proposed is still a patch bolted onto a system that departments are already using to put handcuffs on people. "Out-of-gallery detection" is an admission that the tool has been confidently wrong all along, and that nobody built in a way to say "I don't know" until researchers started publishing papers about it in 2025 and 2026. Meanwhile the databases keep growing, the galleries keep expanding, and agencies keep treating rank-1 output as probable cause first and math problem second.
Facial recognition and ALPR networks like Flock's are two ends of the same surveillance pipeline — one identifies faces, the other identifies plates, and both get treated as infallible until someone's life is upended by an algorithm's overconfidence. If you want to see how much of that pipeline is already running in your area, check our camera map, and if you're ready to push back locally, our take-action page has a starting point.
This is DeFlock The USA’s original commentary. The video above is the work of ND Prof Kevin W. Bowyer, published on YouTube — full credit to the creator.