Certified Robustness on CIFAR-10 (test)
94.5Accuracy (Standard)Diffusion Calibration
Evaluation results
| Method | Links | |||||||||||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Diffusion Calibrationσ=0.25, Backbone=ViT-B/16@2242023.10 | 94.5 | — | 61.9 | — | — | — | — | — | — | — | — | — | 90.3 | 53.5 | — | — | — | — | — | — | — | |
| Carlini et al.σ=0.25, Backbone=ViT-B/16@2242023.10 | 92.8 | — | 59.8 | — | — | — | — | — | — | — | — | — | 89.5 | 52.7 | — | — | — | — | — | — | — | |
| Gaussianσ=0.25, Backbone=ViT-B/16@2242023.10 | 92.7 | — | 55.3 | — | — | — | — | — | — | — | — | — | 87 | 47.9 | — | — | — | — | — | — | — | |
| Cascaded Randomized Smoothingσ=0.25-0.50, Backbone=ViT-B/16@2242023.10 | 92.5 | — | 53.3 | — | 32.9 | — | 14.9 | — | — | — | — | — | 85.1 | 68 | — | — | — | — | — | — | — | |
| Cascaded Randomized Smoothingσ=0.25-1.00, Backbone=ViT-B/16@2242023.10 | 90.1 | — | 40.5 | — | 25 | — | 16.2 | — | 8.8 | — | — | — | 79.6 | 64.5 | — | — | — | — | — | — | — | |
| Consistencyσ=0.25, Backbone=ViT-B/16@2242023.10 | 89.5 | — | 60.7 | — | — | — | — | — | — | — | — | — | 84.9 | 51.7 | — | — | — | — | — | — | — | |
| Carlini et al.σ=0.50, Backbone=ViT-B/16@2242023.10 | 88.8 | — | 50.6 | — | 31 | — | 14.3 | — | — | — | — | — | 75.3 | 64.6 | — | — | — | — | — | — | — | |
| Diffusion Calibrationσ=0.50, Backbone=ViT-B/16@2242023.10 | 88.6 | — | 53.3 | — | 32.5 | — | 15.3 | — | — | — | — | — | 75.1 | 66.4 | — | — | — | — | — | — | — | |
| Gaussianσ=0.50, Backbone=ViT-B/16@2242023.10 | 87.4 | — | 45.9 | — | 24.4 | — | 6 | — | — | — | — | — | 69.9 | 54.3 | — | — | — | — | — | — | — | |
| Carlini et al.σ=1.00, Backbone=ViT-B/16@2242023.10 | 87.3 | — | 32.5 | — | 21.9 | — | 11.8 | — | 6.4 | — | — | — | 48.6 | 53 | — | — | — | — | — | — | — | |
| Diffusion Calibrationσ=1.00, Backbone=ViT-B/16@2242023.10 | 83.3 | — | 36.7 | — | 24.9 | — | 16.1 | — | 8.8 | — | — | — | 52.6 | 57.7 | — | — | — | — | — | — | — | |
| Consistencyσ=0.50, Backbone=ViT-B/16@2242023.10 | 80.7 | — | 49.2 | — | 31.2 | — | 13.1 | — | — | — | — | — | 67.7 | 61.7 | — | — | — | — | — | — | — | |
| Gaussianσ=1.00, Backbone=ViT-B/16@2242023.10 | 75.6 | — | 27.9 | — | 16.9 | — | 9 | — | 4 | — | — | — | 42.2 | 42.6 | — | — | — | — | — | — | — | |
| LipConvnet-30Conv. Type=LOT, CReg loss=true, HH activation=true, Semi-supervised=true2022.10 | 72.08 | — | — | — | — | — | — | — | — | — | — | — | — | — | 62.84 | 51.99 | 41.94 | — | — | — | — | |
| LipConvnet-15Conv. Type=LOT, CReg loss=true, HH activation=true, Semi-supervised=true2022.10 | 71.89 | — | — | — | — | — | — | — | — | — | — | — | — | — | 62.37 | 52.24 | 42.71 | — | — | — | — | |
| LipConvnet-20Conv. Type=LOT, CReg loss=true, HH activation=true, Semi-supervised=true2022.10 | 71.86 | — | — | — | — | — | — | — | — | — | — | — | — | — | 62.86 | 52.24 | 42.39 | — | — | — | — | |
| LipConvnet-10Conv. Type=LOT, CReg loss=true, HH activation=true, Semi-supervised=true2022.10 | 71.82 | — | — | — | — | — | — | — | — | — | — | — | — | — | 62.6 | 51.76 | 42.31 | — | — | — | — | |
| LipConvnet-25Conv. Type=LOT, CReg loss=true, HH activation=true, Semi-supervised=true2022.10 | 71.64 | — | — | — | — | — | — | — | — | — | — | — | — | — | 62.67 | 52 | 42.34 | — | — | — | — | |
| LipConvnet-5Conv. Type=LOT, CReg loss=true, HH activation=true, Semi-supervised=true2022.10 | 71.52 | — | — | — | — | — | — | — | — | — | — | — | — | — | 61.25 | 50.66 | 40.31 | — | — | — | — | |
| LipConvnet-15Conv. Type=SOC, CReg loss=true, HH activation=true, Semi-supervised=true2022.10 | 71.15 | — | — | — | — | — | — | — | — | — | — | — | — | — | 61.65 | 51.78 | 42.53 | — | — | — | — | |
| LipConvnet-10Conv. Type=SOC, CReg loss=true, HH activation=true, Semi-supervised=true2022.10 | 71.1 | — | — | — | — | — | — | — | — | — | — | — | — | — | 60.81 | 50.61 | 41.03 | — | — | — | — | |
| LipConvnet-20Conv. Type=SOC, CReg loss=true, HH activation=true, Semi-supervised=true2022.10 | 70.95 | — | — | — | — | — | — | — | — | — | — | — | — | — | 61.72 | 51.78 | 42.01 | — | — | — | — | |
| LipConvnet-25Conv. Type=SOC, CReg loss=true, HH activation=true, Semi-supervised=true2022.10 | 70.58 | — | — | — | — | — | — | — | — | — | — | — | — | — | 61.4 | 51.46 | 42.05 | — | — | — | — | |
| LipConvnet-30Conv. Type=SOC, CReg loss=true, HH activation=true, Semi-supervised=true2022.10 | 70.37 | — | — | — | — | — | — | — | — | — | — | — | — | — | 60.74 | 51.23 | 41.81 | — | — | — | — | |
| LipConvnet-5Conv. Type=SOC, CReg loss=true, HH activation=true, Semi-supervised=true2022.10 | 69.67 | — | — | — | — | — | — | — | — | — | — | — | — | — | 59.28 | 48.02 | 38.3 | — | — | — | — | |
| Consistencyσ=1.00, Backbone=ViT-B/16@2242023.10 | 65.8 | — | 31.1 | — | 22.3 | — | 14.6 | — | 8.7 | — | — | — | 43.1 | 51.3 | — | — | — | — | — | — | — | |
| Baseline 12026.06 | — | — | — | — | — | — | — | — | — | — | — | — | 70.88 | — | — | — | — | — | 41.63 | 0.08 | 16.99 | |
| Baseline 22026.06 | — | — | — | — | — | — | — | — | — | — | — | — | 70.87 | — | — | — | — | — | 42.64 | 0.07 | 14.92 | |
| Baseline 32026.06 | — | — | — | — | — | — | — | — | — | — | — | — | 70.87 | — | — | — | — | — | 41.61 | 0.08 | 17.04 | |
| Baseline 42026.06 | — | — | — | — | — | — | — | — | — | — | — | — | 72.79 | — | — | — | — | — | 84.74 | 0.11 | 67.58 | |
| Cohen et al.Backbone=ResNet-110, Randomized Smoothing=Gaussian2020.02 | — | 60 | 43 | 34 | 23 | 17 | 14 | 12 | 10 | 8 | 6 | 4 | — | — | — | — | — | — | — | — | — | |
| Consistencysigma=1.002021.11 | — | 42.2 | 38.1 | 34.3 | 30 | 26.3 | 22.9 | 19.7 | — | 13.8 | — | — | 46.3 | 75.6 | — | — | — | — | — | — | — | |
| CURE-Scratchepsilon_2=0.252024.10 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | 71.2 | — | — | — | |
| CURE-Scratchepsilon_2=0.52024.10 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | 56.6 | — | — | — | |
| Gaussiansigma=1.002021.11 | — | 39.2 | 34 | 27.8 | 21.6 | 17.4 | 14 | 11.8 | 10 | 7.6 | — | — | 47.2 | 54.2 | — | — | — | — | — | — | — | |
| Hu et al. (2023)epsilon_2=0.252024.10 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | 69.5 | — | — | — | |
| Hu et al. (2023)epsilon_2=0.52024.10 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | 52.2 | — | — | — | |
| MACERsigma=1.00, official pre-trained=true2021.11 | — | 38.5 | 35.2 | 32.3 | 29.3 | 26.4 | 23.4 | 20.2 | 17.4 | 14.5 | — | — | 41.4 | 74.4 | — | — | — | — | — | — | — | |
| MaxUpBackbone=ResNet-110, Randomized Smoothing=Gaussian2020.02 | — | 74 | 57 | 49 | 40 | 35 | 31 | 27 | 22 | 19 | 17 | 15 | — | — | — | — | — | — | — | — | — | |
| RRISE2026.06 | — | — | — | — | — | — | — | — | — | — | — | — | 70.26 | — | — | — | — | — | 43.68 | 0.08 | 16.1 | |
| Salman et al.Backbone=ResNet-110, Randomized Smoothing=Gaussian (PGD-based)2020.02 | — | 74 | 57 | 48 | 38 | 33 | 29 | 25 | 19 | 17 | 14 | 12 | — | — | — | — | — | — | — | — | — | |
| SmoothAdvsigma=1.00, official pre-trained=true2021.11 | — | 44.9 | 39 | 33.6 | 28.5 | 23.7 | 19.4 | 15.4 | 12 | 8.7 | — | — | 50.8 | 66 | — | — | — | — | — | — | — | |
| SmoothMixsigma=1.002021.11 | — | 42.5 | 37.5 | 32.9 | 28.7 | 24.9 | 21.3 | 18.3 | 15.5 | 12.6 | — | — | 47.1 | 72.5 | — | — | — | — | — | — | — | |
| SmoothMix + One-step adversarysigma=1.002021.11 | — | 41.5 | 37.5 | 33.8 | 30.2 | 26.7 | 23.4 | 20.2 | 17.2 | 14.7 | — | — | 45.1 | 77.3 | — | — | — | — | — | — | — | |
| Stability trainingsigma=1.002021.11 | — | 38.9 | 32.8 | 27 | 23.1 | 19.1 | 15.4 | 11.3 | 7.8 | 5.7 | — | — | 43.5 | 52.6 | — | — | — | — | — | — | — |