Adversarial Attack on ImageNet (Perturbation Metrics)
52.08ParsimonSalman ResNet-18
Evaluation Results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| Salman ResNet-18source=Salman, Architecture=ResNet-18, Iterations=1002025.11 | 52.08 | 51.5 | 52.58 | 52.06 | 56.06 | 3.48 | |
| Wong ResNet-50source=Wong[29], Architecture=ResNet-50, Iterations=1002025.11 | 48.32 | 49.1 | 50.86 | 49.48 | 53.86 | 3 | |
| Engstrom ResNet-50source=Engstrom[15], Architecture=ResNet-50, Iterations=1002025.11 | 42.4 | 41.68 | 42.92 | 42.08 | 45.26 | 2.34 | |
| Salman ResNet-50source=Salman[23], Architecture=ResNet-50, Iterations=1002025.11 | 41.24 | 40.42 | 41.98 | 41.06 | 44.44 | 2.46 | |
| Salman WideResNet-50-2source=Salman, Architecture=WideResNet-50-2, Iterations=1002025.11 | 36.84 | 35.64 | 37.82 | 36.54 | 39.84 | 2.02 | |
| PyTorch ResNet-50source=PyTorch, Architecture=ResNet-50, Iterations=1002025.11 | 33.92 | 47.56 | 50.08 | 38.8 | 47.52 | -2.52 | |
| Debenedetti XCiT-S12source=Debenedetti[12], Architecture=XCiT-S12, Iterations=1002025.11 | 31.72 | 30.66 | 32.36 | 31.64 | 33.86 | 1.5 | |
| Singh ViT-S+ConvStemsource=Singh[25], Architecture=ViT-S+ConvStem, Iterations=1002025.11 | 31.16 | 30.1 | 31.4 | 30.92 | 33.48 | 2.08 | |
| Singh ConvNeXt-T+ConvStemsource=Singh, Architecture=ConvNeXt-T+ConvStem, Iterations=1002025.11 | 30.52 | 29.76 | 30.64 | 30.04 | 32.78 | 2.14 | |
| Debenedetti XCiT-M12source=Debenedetti, Architecture=XCiT-M12, Iterations=1002025.11 | 30.36 | 29.38 | 31.06 | 30.14 | 32.84 | 1.78 | |
| Debenedetti XCiT-L12source=Debenedetti, Architecture=XCiT-L12, Iterations=1002025.11 | 30.12 | 29.58 | 30.66 | 29.94 | 32.32 | 1.66 | |
| Singh ConvNeXt-S+ConvStemsource=Singh, Architecture=ConvNeXt-S+ConvStem, Iterations=1002025.11 | 29.26 | 28.46 | 29.72 | 28.98 | 31.34 | 1.62 | |
| Singh ViT-B+ConvStemsource=Singh, Architecture=ViT-B+ConvStem, Iterations=1002025.11 | 27.12 | 26.4 | 27.56 | 26.68 | 29.22 | 1.66 | |
| Singh ConvNeXt-B+ConvStemsource=Singh, Architecture=ConvNeXt-B+ConvStem, Iterations=1002025.11 | 26.9 | 26.2 | 27.38 | 26.82 | 28.86 | 1.48 | |
| Liu Swin-Bsource=Liu, Architecture=Swin-B, Iterations=1002025.11 | 26.86 | 26.06 | 27.2 | 26.74 | 28.88 | 1.68 | |
| Liu ConvNeXt-Bsource=Liu[18], Architecture=ConvNeXt-B, Iterations=1002025.11 | 26.48 | 25.88 | 26.84 | 26.44 | 28.36 | 1.52 | |
| Singh ConvNeXt-L+ConvStemsource=Singh, Architecture=ConvNeXt-L+ConvStem, Iterations=1002025.11 | 25.36 | 24.82 | 25.94 | 25.34 | 26.94 | 1 | |
| Liu ConvNeXt-Lsource=Liu, Architecture=ConvNeXt-L, Iterations=1002025.11 | 25.08 | 24.26 | 25.78 | 24.9 | 26.88 | 1.1 | |
| Liu Swin-Lsource=Liu, Architecture=Swin-L, Iterations=1002025.11 | 24.16 | 23.36 | 24.62 | 23.8 | 26.06 | 1.44 |