Adversarial ML Lab

FGSM / PGD / C&W attacks + Randomized Smoothing · CIFAR-10 · measured SmallCNN · Source repo ↗
Measured benchmark data from results/cifar10_smallcnn_real.json. On the committed CPU SmallCNN run, clean CIFAR-10 accuracy is 71.82%; PGD-20 at ε=8/255 reduces robust accuracy to 0.00%. This is a measured small-model benchmark, not the synthetic ResNet-18 projection.
Run live attack evaluation
Live accuracy: —
0%
Clean accuracy
CIFAR-10 SmallCNN · 10,000 clean samples
0%
PGD accuracy
PGD-20 · ε=8/255 · 1,024 samples
0%
FGSM accuracy
ε=8/255 · 1,024 samples
0%
C&W accuracy
L2 · c=1.0 · 1,024 samples
Accuracy under attack — CIFAR-10
Attack vs defense summary
Clean (baseline)
71.82%
FGSM ε=8/255
3.32%
PGD-20 ε=8/255
0.00%
C&W L2
4.20%