Image Classification on ImageNet (Accuracy and Search/Robustness Metrics)
82AccuracyBase model
Evaluation results
| Method | Links | |||||||
|---|---|---|---|---|---|---|---|---|
| Base modelBackbone=DeiT2023.10 | 82 | 6.4 | 0 | 46.7 | 0.8 | 22.3 | 0 | |
| Randomized Feature DefenseBackbone=DeiT, Accuracy drop constraint=≈ 1%2023.10 | 80.8 | 69.7 | 69.1 | 75 | 59.1 | 66.4 | 64.1 | |
| Base modelBackbone=ResNet502023.10 | 80.37 | 3.5 | 0.2 | 36.2 | 4.3 | 6.6 | 0.4 | |
| Randomized Input DefenseBackbone=DeiT, Accuracy drop constraint=≈ 1%2023.10 | 80.1 | 67.7 | 67.2 | 75.8 | 65.9 | 64.4 | 63.6 | |
| Randomized Feature DefenseBackbone=DeiT, Accuracy drop constraint=≈ 2%2023.10 | 79.76 | 69.3 | 69 | 75.1 | 65.3 | 66 | 64.3 | |
| Randomized Feature DefenseBackbone=ResNet50, Accuracy drop constraint=≈ 1%2023.10 | 79.7 | 37 | 36 | 56.7 | 16.8 | 46.3 | 43.4 | |
| Randomized Input DefenseBackbone=DeiT, Accuracy drop constraint=≈ 2%2023.10 | 79.6 | 66.6 | 66 | 75.7 | 67.1 | 64.9 | 64.3 | |
| Randomized Input DefenseBackbone=ResNet50, Accuracy drop constraint=≈ 1%2023.10 | 79.18 | 40.3 | 39.5 | 63.8 | 23.9 | 47.6 | 45.4 | |
| Base modelBackbone=ViT2023.10 | 79.15 | 5.7 | 0 | 45.7 | 7.3 | 5.1 | 0 | |
| CLIP-1+DINO(v2+v3)Setting=Inductive, Test-Time Adaptation (TTA)=false2025.06 | 79.1 | — | — | — | — | — | — | |
| Randomized Input DefenseBackbone=ResNet50, Accuracy drop constraint=≈ 2%2023.10 | 78.46 | 41.1 | 39.8 | 69.4 | 41.5 | 49.3 | 47.2 | |
| Randomized Feature DefenseBackbone=ResNet50, Accuracy drop constraint=≈ 2%2023.10 | 78.43 | 42 | 41.5 | 65.6 | 40.6 | 51.3 | 49.3 | |
| Randomized Input DefenseBackbone=ViT, Accuracy drop constraint=≈ 1%2023.10 | 78.28 | 58.8 | 58.1 | 70.8 | 51.4 | 53.1 | 52.2 | |
| Randomized Feature DefenseBackbone=ViT, Accuracy drop constraint=≈ 1%2023.10 | 78.2 | 60.6 | 60.2 | 69.1 | 47.5 | 54 | 52.9 | |
| COSMIC*Setting=Inductive, Test-Time Adaptation (TTA)=true2025.06 | 78.2 | — | — | — | — | — | — | |
| CLIP-1+DINOv3Setting=Inductive, Test-Time Adaptation (TTA)=false2025.06 | 77.9 | — | — | — | — | — | — | |
| Randomized Feature DefenseBackbone=ViT, Accuracy drop constraint=≈ 2%2023.10 | 77.18 | 63.7 | 62.9 | 72.2 | 58.1 | 57 | 55.3 | |
| Randomized Input DefenseBackbone=ViT, Accuracy drop constraint=≈ 2%2023.10 | 77.09 | 61.3 | 60.9 | 70.6 | 59.2 | 53.7 | 52.7 | |
| CLIP-1+DINOv2Setting=Inductive, Test-Time Adaptation (TTA)=false2025.06 | 77 | — | — | — | — | — | — | |
| Black-box2026.03 | 76.15 | — | — | — | — | — | — | |
| MorphSNNT=42026.03 | 75.04 | — | — | — | — | — | — | |
| SpikingformerT=42026.03 | 74.79 | — | — | — | — | — | — | |
| S-TransformerT=42026.03 | 74.57 | — | — | — | — | — | — | |
| Base modelBackbone=VGG192023.10 | 74.21 | 0.1 | 0 | 19.6 | 0 | 0.4 | 0 | |
| M-CBMSparsity=NCC=avg2026.03 | 73.64 | — | — | — | — | — | — | |
| SpikformerT=42026.03 | 73.38 | — | — | — | — | — | — | |
| Randomized Input DefenseBackbone=VGG19, Accuracy drop constraint=≈ 1%2023.10 | 73.24 | 7.7 | 6.9 | 32.1 | 1.5 | 18.3 | 17 | |
| Randomized Feature DefenseBackbone=VGG19, Accuracy drop constraint=≈ 1%2023.10 | 72.66 | 22.4 | 21.6 | 50.1 | 18.5 | 34.6 | 32.9 | |
| DMN*Setting=Inductive, Test-Time Adaptation (TTA)=true2025.06 | 72.2 | — | — | — | — | — | — | |
| M-CBMSparsity=NCC=52026.03 | 72.18 | — | — | — | — | — | — | |
| DPE*Setting=Inductive, Test-Time Adaptation (TTA)=true2025.06 | 71.9 | — | — | — | — | — | — | |
| Randomized Input DefenseBackbone=VGG19, Accuracy drop constraint=≈ 2%2023.10 | 71.43 | 18.7 | 17.8 | 47.4 | 11.5 | 28.3 | 27.1 | |
| Randomized Feature DefenseBackbone=VGG19, Accuracy drop constraint=≈ 2%2023.10 | 71.21 | 23.3 | 22.2 | 55.1 | 28.4 | 36.5 | 32.8 | |
| ZeroSetting=Inductive, Test-Time Adaptation (TTA)=false2025.06 | 71.2 | — | — | — | — | — | — | |
| TIPPLE*Setting=Inductive, Test-Time Adaptation (TTA)=true2025.06 | 71 | — | — | — | — | — | — | |
| MTA*Setting=Inductive, Test-Time Adaptation (TTA)=true2025.06 | 70.1 | — | — | — | — | — | — | |
| DN-CBMViTSparsity=NCC=avg2026.03 | 69.98 | — | — | — | — | — | — | |
| TDA*Setting=Inductive, Test-Time Adaptation (TTA)=true2025.06 | 69.5 | — | — | — | — | — | — | |
| LF-CBMSparsity=NCC=avg2026.03 | 69.08 | — | — | — | — | — | — | |
| CLIP-1+CLIP-2Setting=Inductive, Test-Time Adaptation (TTA)=false2025.06 | 68.9 | — | — | — | — | — | — | |
| CLIP-1Setting=Inductive, Test-Time Adaptation (TTA)=false2025.06 | 66.6 | — | — | — | — | — | — | |
| LF-CBMSparsity=NCC=52026.03 | 62.2 | — | — | — | — | — | — | |
| DN-CBMViTSparsity=NCC=52026.03 | 60.23 | — | — | — | — | — | — | |
| CLIP-2Setting=Inductive, Test-Time Adaptation (TTA)=false2025.06 | 58.2 | — | — | — | — | — | — | |
| DN-CBMRNSparsity=NCC=avg2026.03 | 57.24 | — | — | — | — | — | — | |
| DN-CBMRNSparsity=NCC=52026.03 | 46.71 | — | — | — | — | — | — |