LPID confines perturbations to the face region and optimizes them through a differentiable crop-and-resize model, holding attacker accuracy below 10% on unseen identities where prior unlearnable-example methods collapse to 37–74%.
Title resolution pending
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CV 1years
2026 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Unlearnable Faces: Privacy Protection Surviving Extraction Pipeline
LPID confines perturbations to the face region and optimizes them through a differentiable crop-and-resize model, holding attacker accuracy below 10% on unseen identities where prior unlearnable-example methods collapse to 37–74%.