Skip to main content

Table 7 Adversarial distortion

From: Machine learning through cryptographic glasses: combating adversarial attacks by key-based diversified aggregation

AttackMedian 2-normMean 2-norm
MNIST
C&W25.28e −035.52e −03
C&W01.56e −021.61e −02
C&W1.24e −021.29e −02
Fashion-MNIST
C&W22.30e −045.31e −04
C&W04.35e −034.86e −03
C&W4.43e −045.43e −04
CIFAR-10
C&W27.80e −051.19e −04
C&W02.48e −034.55e −03
C&W1.73e −042.13e −04
ResNet18 (CIFAR-10)
PGD1.00e −041.37e −04
Vanilla OnePixelp=15.25e −042.76e −03
Vanilla OnePixelp=31.24e −033.51e −03
Vanilla OnePixelp=51.86e −034.18e −03
Multi-channel model OnePixelp=13.22e −042.42e −03
Multi-channel model OnePixelp=31.33e −033.61e −03
Multi-channel model OnePixelp=52.05e −034.33e −03
VGG16 (CIFAR-10)
PGD9.99e −051.50e −04
Vanilla OnePixelp=15.86e −042.78e −03
Vanilla OnePixelp=31.37e −033.69e −03
Vanilla OnePixelp=52.02e −034.26e −03
Multi-channel model OnePixelp=13.43e −042.25e −03
Multi-channel model OnePixelp=31.27e −033.64e −03
Multi-channel model OnePixelp=51.91e −034.28e −03