Image Classification on Tiny-ImageNet (val) (Adversarial Robustness)
69Clean AccuracyMPM
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| MPMBackbone=ResNeXt-50, Resolution=224x224, Param.=23.4M, FLOPs=4.3G2026.01 | 69 | 48.9 | 28.5 | |
| ours (vanilla)Backbone=ResNeXt-50, Resolution=224x224, Param.=23.4M, FLOPs=4.3G2026.01 | 67.9 | 48 | 27.6 | |
| Wang et al.Backbone=WideResNet-28-10, Resolution=64x64, Param.=36.6M, FLOPs=21.0G, Note=Original results with label leakage2026.01 | 65.2 | 48.3 | 31.3 | |
| Gowal et al.Backbone=WideResNet-28-10, Resolution=64x64, Param.=36.6M, FLOPs=21.0G, Note=Original results with label leakage2026.01 | 61 | — | 26.7 | |
| Wang et al.Backbone=WideResNet-28-10, Resolution=64x64, Param.=36.6M, FLOPs=21.0G, Note=Reassessed performance2026.01 | 57.6 | 38.4 | 28.4 | |
| Dilation (ours)Architecture=Dilation (ours), Training Method=FastAT, Training GPU Days=0.222021.08 | 46.22 | 23.9 | — | |
| ResNet-50Architecture=Backbone, Training Method=FastAT, Training GPU Days=0.122021.08 | 45.92 | 23.53 | — | |
| Dilation (ours)Architecture=Dilation (ours), Training Method=FreeAT-4, Training GPU Days=0.122021.08 | 44.68 | 24.66 | — | |
| Dilation (ours)Architecture=Dilation (ours), Training Method=PGD-4, Training GPU Days=0.452021.08 | 44.37 | 24.33 | — | |
| ResNet-50Architecture=Backbone, Training Method=PGD-4, Training GPU Days=0.192021.08 | 43.23 | 24.13 | — | |
| ResNet-50Architecture=Backbone, Training Method=FreeAT-4, Training GPU Days=0.052021.08 | 42.73 | 24.1 | — |