Abstract:
"A more focused and efficient assault is promoted by the use of Grad-CAM++, which guarantees
that the adversarial noise is aligned with significant picture features. The suggested strategy is
shown to be effective in producing adversarial cases through an experimental assessment. The
concepts for the creation of model resistance strategies for future researchers are also
encouraged by this research, which advances adversarial assault tactics.
The target model, a face recognition model that bridges the FaceNet backbone structure with
InceptionResNet-v1, was used for the experiment. For 500 subsets of the CASIA-WebFace
dataset, the XDONoise PGD variant was able to archive 99.0% attack success rate with less
magnitude change better than original PGD "
