NinjaDesc: Content-Concealing Visual Descriptors via Adversarial Learning