Experimental Facebook face-recogniton software can discern identity with 83% accuracy even when faces are hidden
Facebook can recognise you in photos even if you're not looking — Thanks to the latest advances in computer vision, we now have machines that can pick you out of a line-up.
Context & Ripple Effects
This 2015 demonstration is the opening move in a fight that Facebook itself kept escalating. A year after the hidden-face result, it released its object-detection stack to developers via the DeepMask, SharpMask, and MultiPathNet open-source drop, spreading the underlying vision techniques beyond its own walls.
The reaction side of the ledger arrived quickly: researchers showed printed paper eyeglass frames could fool many recognizers, and by 2019 Facebook AI Research was building its own live-video face-modifying system to defeat state-of-the-art recognition — the same lab lineage that produced the 83% hidden-face result now supplying the countermeasure.
First-order effects
- Anyone appearing in photos without looking at the camera becomes identifiable by Facebook's system, extending the reach of automatic tagging from posed shots to candid crowds.
Second-order effects
- Adversarial countermeasures become a research field in their own right: the eyeglasses attack and later the University of Chicago's Fawkes pixel-level photo disguises exist only because recognition this strong creates demand for evasion.
Third-order effects
- Identity verification splits into an arms race where the strongest recognizer's owner also fields the defense — Facebook ends up on both sides, and the contest shifts from whether faces can be recognized to who controls the cloaking layer around them.
The trend: Facial recognition is settling into a co-evolutionary arms race in which each accuracy advance provokes a matching obfuscation technique, with platform labs funding both sides.