Robust vs Vanilla Image Classifiers

Robust vs. Vanilla Image Classifiers

Standard accuracy of adversarially-trained models (from RobustBench) compared against vanilla / standard models (from Papers with Code). Adversarial-training papers rarely report the vanilla SOTA they compare against, and the underlying vanilla model is usually unknown — so several comparison references are shown (best / median / mean / top-10 statistics of all PwC entries, the same restricted to vanilla (no extra data) models, and a hand-curated architecture-matched baseline whose standard accuracies are taken from the torchvision model zoo for ImageNet, CIFAR-10/100 data from kuangliu/pytorch-cifar and weiaicunzai/pytorch-cifar100 and from the original architecture papers for CIFAR-10/100 and ImageNet. Robust accuracy is the standardized AutoAttack value.

Loading data...

Models shown

Comparison reference

Direct match only Hide proxy-matched architectures

Clean accuracy vs. robust accuracy (AutoAttack)

One point per robust model; colour = publication year. Vertical dashed line = selected vanilla reference.

Clean vs. robust vs. vanilla reference (top-N)

Grouped bars per robust model; horizontal line = selected vanilla reference.

Model table

Click a column header to sort. Δ = model value − reference (negative = model is below the reference).

Sources: Robust accuracy & clean accuracy of robust models: RobustBench model_info (ranked by AutoAttack). Vanilla / standard accuracy: Papers with Code API (image-classification leaderboards for CIFAR-10, CIFAR-100, ImageNet-1k). (ImageNet values from the torchvision model zoo, CIFAR-10/100 data from kuangliu/pytorch-cifar and weiaicunzai/pytorch-cifar100 and from the original architecture papers for CIFAR-10/100 and ImageNet. Note that Papers with Code's leaderboards are now curated/trimmed by Hugging Face, so reference statistics reflect the live curated boards rather than an exhaustive historical list.