Geometric Robustness Certification on TinyImageNet VNN-Comp'24 model
18Certified Accuracy (%)Super-DeepG
Evaluation results
| Method | Links | ||
|---|---|---|---|
| Super-DeepGInterval size (°)=0.022026.04 | 18 | 3.73 | |
| Super-DeepGInterval size=0.022026.04 | 18 | 3.73 | |
| FGVInterval size (°)=2 · 10−42026.04 | 10 | 10.9 | |
| Super-DeepGInterval size (°)=0.042026.04 | 10 | 1.46 | |
| FGVInterval size=2 × 10−42026.04 | 10 | 10.09 | |
| FGVInterval size (°)=4 · 10−42026.04 | 2 | 5.06 | |
| FGVInterval size (°)=3 · 10−42026.04 | 2 | 6.78 | |
| FGVInterval size=4 × 10−42026.04 | 2 | 5.06 | |
| FGVInterval size=3 × 10−42026.04 | 2 | 6.78 | |
| FGVInterval size=5 × 10−42026.04 | 1 | 4.05 |