Wrong Inputs, Wrong Thresholds: Re-evaluating Adversarial Face Privacy
Wrong Input, Wrong Threshold: Re-evaluating Adversarial Facial Privacy
The paper re-evaluates published adversarial face privacy methods and finds that their reported black-box protection rates above 90% stem from flawed evaluation protocols. Rescored against the detection, alignment and cropping front end of real deployed systems, standard transfer attacks lose 56% of their evasion protection and budget-allocation attacks lose 7%; the original metric saturates, so the authors propose a pairwise cosine displacement in which thresholds cancel algebraically, measuring input defects 2.2x and 10.7x larger on three unseen recognizers and reversing the L-infinity transfer ranking at equal budget. The defect's root cause is resampling mismatch rather than loss of out-of-crop perturbation, and it is largely eliminated by optimization through a differentiable alignment operator, improving by 24 to 58 percentage points on unseen recognizers and one commercial system, with residual effects remaining on one. Four threshold constants carried over from 2021 keep the false-accept rate near the nominal 1% on aligned crops but relax it to 2.3-5.4% on full images, and bias the sign in opposite directions on evasion and impersonation tasks.