Image Classification on ImageNet RobustBench (val)
76.62Clean AccuracyViT-B
Evaluation results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| ViT-BParams (M)=86.6, FLOPs (G)=17.6, Source=[44], Adversarial Training Epochs=300, Adv. Steps=2, Model Size=Large2023.03 | 76.62 | 53.5 | — | — | — | |
| ViT-B + ConvStemParams (M)=87.1, FLOPs (G)=17.9, Source=ours, Adversarial Training Epochs=250, Adv. Steps=2, Model Size=Large2023.03 | 76.3 | 54.66 | 56.3 | 32.06 | — | |
| Swin-BParams (M)=87.7, FLOPs (G)=15.5, Source=[32]*, Adversarial Training Epochs=300, Adv. Steps=3, Model Size=Large2023.03 | 76.16 | 56.16 | 47.86 | 23.91 | — | |
| ConvNeXt-BParams (M)=88.6, FLOPs (G)=15.4, Source=[32]*, Adversarial Training Epochs=300, Adv. Steps=3, Model Size=Large2023.03 | 76.02 | 55.82 | 44.68 | 21.23 | — | |
| ConvNeXt-B + ConvStemParams (M)=88.8, FLOPs (G)=16, Source=ours, Adversarial Training Epochs=250, Adv. Steps=2, Model Size=Large2023.03 | 75.9 | 56.14 | 49.12 | 23.34 | — | |
| ConvNeXt-BParams (M)=88.6, FLOPs (G)=15.4, Source=ours, Adversarial Training Epochs=50, Adv. Steps=2, Model Size=Large2023.03 | 75.62 | 54.34 | 48.52 | 23.7 | — | |
| ConvNeXt-B + ConvStemParams (M)=88.8, FLOPs (G)=16, Source=ours, Adversarial Training Epochs=50, Adv. Steps=2, Model Size=Large2023.03 | 75.32 | 54.38 | 50.06 | 24.76 | — | |
| ConvNeXt-B + ConvStemParams (M)=88.8, FLOPs (G)=16, Source=ours, Adversarial Training Epochs=250, Adv. Steps=3, Model Size=Large2023.03 | 75.18 | 56.28 | 49.4 | 23.6 | — | |
| Swin-BParams (M)=87.7, FLOPs (G)=15.5, Source=[37], Adversarial Training Epochs=90, Adv. Steps=3, Model Size=Large2023.03 | 74.76 | 48.1 | 44.42 | 18.04 | — | |
| ViT-B + ConvStemParams (M)=87.1, FLOPs (G)=17.9, Source=ours, Adversarial Training Epochs=50, Adv. Steps=2, Model Size=Large2023.03 | 74.38 | 52.58 | 54.38 | 31.2 | — | |
| ConvNeXt-SParams (M)=50.1, FLOPs (G)=8.7, Source=ours, Adversarial Training Epochs=50, Adv. Steps=2, Model Size=Medium2023.03 | 74.1 | 52.32 | 43.84 | 19.52 | — | |
| ConvNeXt-S + ConvStemParams (M)=50.3, FLOPs (G)=8.8, Source=ours, Adversarial Training Epochs=50, Adv. Steps=2, Model Size=Medium2023.03 | 74.1 | 52.42 | 50.88 | 25.64 | — | |
| XCIT-M12Params (M)=46, FLOPs (G)=8.5, Source=[14], Adversarial Training Epochs=110, Adv. Steps=1, Model Size=Medium2023.03 | 74.04 | 45.24 | 48.18 | 22.72 | — | |
| XCIT-L12Params (M)=104, FLOPs (G)=19, Source=[14], Adversarial Training Epochs=110, Adv. Steps=1, Model Size=Large2023.03 | 73.76 | 47.6 | 49.38 | 23.74 | — | |
| RobArch-LParams (M)=104, FLOPs (G)=25.7, Source=[42], Adversarial Training Epochs=100, Adv. Steps=3, Model Size=Large2023.03 | 73.46 | 48.92 | 39.48 | 14.74 | — | |
| ViT-BParams (M)=86.6, FLOPs (G)=17.6, Source=ours, Adversarial Training Epochs=50, Adv. Steps=2, Model Size=Large2023.03 | 73.32 | 50.02 | 52.14 | 33.12 | — | |
| ConvNeXt-T + ConvStemParams (M)=28.6, FLOPs (G)=4.6, Source=ours, Adversarial Training Epochs=300, Adv. Steps=2, Model Size=Small2023.03 | 72.72 | 49.46 | 48.42 | 24.52 | — | |
| ConvNeXt-T + ConvStemParams (M)=28.6, FLOPs (G)=4.6, Source=ours, Adversarial Training Epochs=300, Adv. Steps=3, Model Size=Small2023.03 | 72.7 | 50.16 | 49 | 24.16 | — | |
| ViT-S + ConvStemParams (M)=22.8, FLOPs (G)=5, Source=ours, Adversarial Training Epochs=300, Adv. Steps=2, Model Size=Small2023.03 | 72.56 | 48.08 | 50.4 | 26.68 | — | |
| ConvNeXt-TParams (M)=28.6, FLOPs (G)=4.5, Source=ours, Adversarial Training Epochs=300, Adv. Steps=2, Model Size=Small2023.03 | 72.4 | 48.6 | 38.02 | 14.88 | — | |
| ViT-M + ConvStemParams (M)=39.5, FLOPs (G)=8.4, Source=ours, Adversarial Training Epochs=50, Adv. Steps=2, Model Size=Medium2023.03 | 72.4 | 48.8 | 50.56 | 28.12 | — | |
| XCIT-S12Params (M)=26, FLOPs (G)=4.8, Source=[14], Adversarial Training Epochs=110, Adv. Steps=1, Model Size=Small2023.03 | 72.34 | 41.78 | 46.2 | 22.72 | — | |
| ViT-MParams (M)=38.8, FLOPs (G)=8, Source=ours, Adversarial Training Epochs=50, Adv. Steps=2, Model Size=Medium2023.03 | 71.72 | 47.24 | 49.02 | 29.2 | — | |
| ConvNeXt-TParams (M)=28.6, FLOPs (G)=4.5, Source=[14], Adversarial Training Epochs=110, Adv. Steps=1, Model Size=Small2023.03 | 71.6 | 44.4 | 45.32 | 21.76 | — | |
| RobArch-SParams (M)=26.1, FLOPs (G)=6.3, Source=[42], Adversarial Training Epochs=110, Adv. Steps=3, Model Size=Small2023.03 | 70.58 | 44.12 | 39.88 | 15.46 | — | |
| ViT-BParams (M)=86.6, FLOPs (G)=17.6, Source=[37], Adversarial Training Epochs=90, Adv. Steps=3, Model Size=Large2023.03 | 70.42 | 43.02 | 47.26 | 27.08 | — | |
| Isotropic-CN-S + ConvStemParams (M)=23, FLOPs (G)=4.7, Source=ours, Adversarial Training Epochs=300, Adv. Steps=2, Model Size=Small2023.03 | 70.02 | 45.9 | 49.24 | 27.84 | — | |
| ResNet-101Params (M)=44.5, FLOPs (G)=7.9, Source=[37], Adversarial Training Epochs=90, Adv. Steps=3, Model Size=Medium2023.03 | 69.52 | 41.02 | 25.62 | 6.56 | — | |
| ViT-SParams (M)=22.1, FLOPs (G)=4.6, Source=ours, Adversarial Training Epochs=300, Adv. Steps=2, Model Size=Small2023.03 | 69.22 | 44.04 | 37.52 | 15.12 | — | |
| Isotropic-CN-SParams (M)=22.3, FLOPs (G)=4.3, Source=ours, Adversarial Training Epochs=300, Adv. Steps=2, Model Size=Small2023.03 | 69.04 | 44.22 | 36.64 | 14.88 | — | |
| Wide-ResNet-50-2Params (M)=68.9, FLOPs (G)=11.4, Source=[46], Adversarial Training Epochs=100, Adv. Steps=3, Model Size=Medium2023.03 | 68.82 | 38.12 | 22.08 | 4.48 | — | |
| ResNet-50Params (M)=25, FLOPs (G)=4.1, Source=[2], Adversarial Training Epochs=100, Adv. Steps=1, Model Size=Small2023.03 | 67.44 | 35.54 | 18.16 | 3.9 | — | |
| ViT-SParams (M)=22.1, FLOPs (G)=4.6, Source=[14], Adversarial Training Epochs=110, Adv. Steps=1, Model Size=Small2023.03 | 66.78 | 37.88 | — | — | — | |
| ViT-SParams (M)=22.1, FLOPs (G)=4.6, Source=[2], Adversarial Training Epochs=100, Adv. Steps=1, Model Size=Small2023.03 | 66.62 | 36.56 | 41.4 | 21.82 | — | |
| ViT-SParams (M)=22.1, FLOPs (G)=4.6, Source=[37], Adversarial Training Epochs=90, Adv. Steps=3, Model Size=Small2023.03 | 65.9 | 39.24 | 32.18 | 10.54 | — | |
| ResNet-50Params (M)=25, FLOPs (G)=4.1, Source=[46], Adversarial Training Epochs=100, Adv. Steps=3, Model Size=Small2023.03 | 65.88 | 33.18 | 18.88 | 3.82 | — | |
| SGDcheckpoint_selection=Best2026.05 | 56.02 | 22.22 | 37.3 | 23.63 | 20.17 | |
| SGDcheckpoint_selection=Last2026.05 | 51.03 | 8.4 | 23.39 | 11.72 | 7.76 | |
| Muoncheckpoint_selection=Last2026.05 | 50.93 | 9.33 | 16.8 | 2.98 | 2.5 | |
| Muoncheckpoint_selection=Best2026.05 | 48.44 | 19.04 | 31.45 | 18.12 | 16.21 | |
| AdamWcheckpoint_selection=Best2026.05 | 42.43 | 7.62 | 17.87 | 2.88 | 2.49 | |
| AdamWcheckpoint_selection=Last2026.05 | 42.09 | 7.28 | 18.41 | 9.18 | 6.35 |