Fig. 4
From: Cancelable templates for secure face verification based on deep learning and random projections
ROC curve for the BiometricNet+ performance at inference time on 30000 matching and 30000 non-matching pairs of original and NbNet pre-images recovered from \(r=128\) embeddings; the curve, which exhibits an almost perfect diagonal behavior (e.g., \(GAR = FAR\)), confirms that an attacker would not be able to use NbNet pre-images to access the system even when the projection matrix and the secure templates are disclosed to the public