Image Classification on CIFAR10 (Robustness Metrics: Square, NES, SignHunt)
98.4Clean AccuracyBase
Evaluation Results
| Method | Links | |||||||
|---|---|---|---|---|---|---|---|---|
| BaseBackbone=DeiT2023.10 | 98.4 | 320 | 0 | 81.9 | 34.2 | 7.9 | 20 | |
| BaseBackbone=ViT2023.10 | 97.86 | 510 | 0 | 84.8 | 43.6 | 6.1 | 0 | |
| BaseBackbone=ResNet502023.10 | 97.66 | 80 | 0.1 | 71.7 | 21.7 | 3.7 | 20 | |
| InputBackbone=DeiT, Allowed Clean Accuracy Drop=≈ 2%2023.10 | 96.59 | 6,690 | 67.6 | 95.2 | 90 | 40.2 | 3,920 | |
| FeatureBackbone=DeiT, Allowed Clean Accuracy Drop=≈ 2%2023.10 | 96.29 | 6,910 | 67.9 | 94.1 | 88.3 | 45.7 | 4,340 | |
| BaseBackbone=VGG192023.10 | 96.28 | 60 | 0.1 | 68.8 | 16.6 | 3.2 | 30 | |
| InputBackbone=ResNet50, Allowed Clean Accuracy Drop=≈ 2%2023.10 | 95.98 | 5,050 | 48.8 | 93.1 | 85.4 | 26.8 | 2,600 | |
| FeatureBackbone=ViT, Allowed Clean Accuracy Drop=≈ 2%2023.10 | 95.96 | 6,390 | 62.7 | 93.7 | 85.6 | 42.5 | 4,070 | |
| FeatureBackbone=ResNet50, Allowed Clean Accuracy Drop=≈ 2%2023.10 | 95.95 | 5,490 | 52.8 | 93.2 | 86.2 | 32.5 | 3,060 | |
| InputBackbone=ViT, Allowed Clean Accuracy Drop=≈ 2%2023.10 | 95.8 | 6,300 | 61.2 | 93.5 | 87 | 34.8 | 3,330 | |
| FeatureBackbone=VGG19, Allowed Clean Accuracy Drop=≈ 2%2023.10 | 94.93 | 6,100 | 58.4 | 92.2 | 77.9 | 43.2 | 4,240 | |
| InputBackbone=VGG19, Allowed Clean Accuracy Drop=≈ 2%2023.10 | 94.92 | 3,060 | 27.1 | 89.5 | 58 | 22.7 | 2,180 | |
| FeatureBackbone=DeiT, Allowed Clean Accuracy Drop=≈ 4%2023.10 | 94.91 | 6,890 | 66.1 | 93.5 | 87.6 | 43.6 | 4,040 | |
| InputBackbone=DeiT, Allowed Clean Accuracy Drop=≈ 4%2023.10 | 94.81 | 7,060 | 68.8 | 92.6 | 87.7 | 40.3 | 3,850 | |
| FeatureBackbone=VGG19, Allowed Clean Accuracy Drop=≈ 4%2023.10 | 93.58 | 6,420 | 62.8 | 91.2 | 80.1 | 49.2 | 4,690 | |
| InputBackbone=VGG19, Allowed Clean Accuracy Drop=≈ 4%2023.10 | 93.52 | 4,220 | 39.8 | 90.3 | 68.4 | 27.5 | 2,680 | |
| FeatureBackbone=ResNet50, Allowed Clean Accuracy Drop=≈ 4%2023.10 | 93.48 | 5,670 | 53.4 | 89.9 | 83.9 | 37.1 | 3,570 | |
| InputBackbone=ResNet50, Allowed Clean Accuracy Drop=≈ 4%2023.10 | 93.42 | 5,640 | 54.8 | 90 | 85 | 31.1 | 2,980 | |
| InputBackbone=ViT, Allowed Clean Accuracy Drop=≈ 4%2023.10 | 93.4 | 6,260 | 61.1 | 89.7 | 85.5 | 33.4 | 3,220 | |
| FeatureBackbone=ViT, Allowed Clean Accuracy Drop=≈ 4%2023.10 | 93.39 | 6,620 | 65.6 | 92.9 | 85.3 | 44.8 | 4,380 |