Certified Robustness on ImageNet
71.6Certified Accuracy (L1 R=0.5)Uniform, LBS
Evaluation results
| Method | Links | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| Uniform, LBSNoise Distribution=Uniform2026.04 | 71.6 | 66.3 | 61.4 | 54.6 | 52.5 | 49.9 | — | — | — | |
| Laplace, LBSNoise Distribution=Laplace2026.04 | 70.5 | 63.1 | 57 | 52.8 | 50.7 | 48.1 | — | — | — | |
| Uniform (w/), Yang et al. (2020)Noise Distribution=Uniform, Noise-Augmented Training=true2026.04 | 53.6 | 48.2 | 42.5 | 37.7 | 35.1 | 30.8 | — | — | — | |
| Laplace (w/), Teng et al. (2020)Noise Distribution=Laplace, Noise-Augmented Training=true2026.04 | 47.1 | 39.7 | 31.6 | 26.5 | 23 | 18.2 | — | — | — | |
| Laplace (w/), Lecuyer et al. (2019)Noise Distribution=Laplace, Noise-Augmented Training=true2026.04 | 42.8 | 39.5 | 34.2 | 22.8 | 18.9 | 10.4 | — | — | — | |
| Laplace (w/o), Teng et al. (2020)Noise Distribution=Laplace, Noise-Augmented Training=false2026.04 | 5.8 | 4.2 | 2.9 | 2.1 | 1.4 | 0.7 | — | — | — | |
| Uniform (w/o), Yang et al. (2020)Noise Distribution=Uniform, Noise-Augmented Training=false2026.04 | 5.7 | 2.7 | 2 | 1.6 | 1.1 | 0.7 | — | — | — | |
| Laplace (w/o), Lecuyer et al. (2019)Noise Distribution=Laplace, Noise-Augmented Training=false2026.04 | 5.4 | 3.9 | 3 | 2.2 | 1.8 | 1 | — | — | — | |
| KT PriorSigma=0.25, Sample Budget=10,0002026.06 | — | — | — | — | — | — | 76.2 | 5.18 | 232.77 | |
| KT PriorSigma=0.50, Sample Budget=10,0002026.06 | — | — | — | — | — | — | 73.8 | 4.9 | 263.36 | |
| KT PriorSigma=1.00, Sample Budget=10,0002026.06 | — | — | — | — | — | — | 72.4 | 5.28 | 288.41 | |
| Meta-1-Dynamic-MarginSigma=0.25, Sample Budget=10,0002026.06 | — | — | — | — | — | — | 76.2 | 2.47 | 232.773 | |
| Meta-1-Dynamic-MarginSigma=0.50, Sample Budget=10,0002026.06 | — | — | — | — | — | — | 73.8 | 2.49 | 263.36 | |
| Meta-1-Dynamic-MarginSigma=1.00, Sample Budget=10,0002026.06 | — | — | — | — | — | — | 72.4 | 3.33 | 534.783 |