Image Classification on UCF101 (Acc, Rob)
80.4AccuracyGPUA
Evaluation results
| Method | Links | ||
|---|---|---|---|
| GPUASamples per class=full training set2026.06 | 80.4 | — | |
| GPUA*Samples per class=162026.06 | 78.4 | — | |
| CLBPShots=16, Training Epochs=10, Training Attack (PGD)=2-step (1/255), Evaluation Attack (PGD-100)=100-step (1/255)2026.05 | 76.69 | 71.03 | |
| COSMICInference-time adaptation=true2026.06 | 76.2 | — | |
| StatA2026.06 | 73.5 | — | |
| DMN2026.06 | 72.5 | — | |
| ZLaP2026.06 | 71.5 | — | |
| TIPPLEInference-time adaptation=true2026.06 | 71.2 | — | |
| TDA2026.06 | 70.7 | — | |
| DPE2026.06 | 70.4 | — | |
| ZERO2026.06 | 69.2 | — | |
| AdvVLPShots=16, Training Epochs=10, Training Attack (PGD)=2-step (1/255), Evaluation Attack (PGD-100)=100-step (1/255)2026.05 | 69.1 | 32.8 | |
| AdvMaPLeShots=16, Training Epochs=10, Training Attack (PGD)=2-step (1/255), Evaluation Attack (PGD-100)=100-step (1/255)2026.05 | 68.97 | 31.67 | |
| MTA2026.06 | 68.7 | — | |
| CLBPSetting=Zero-shot, Backbone=ViT-B/16, Attack=PGD-100, Perturbation Budget (epsilon_a)=4/255, Patch Granularity=16x162026.05 | 68.2 | 52.6 | |
| DBDBackbone=ViT-B/16, Attack Protocol=PGD-100, Epsilon (ε)=4/255, Evaluation Setting=Zero-shot2026.06 | 68.2 | 95.9 | |
| TPTSetting=Zero-shot, Backbone=ViT-B/16, Attack=PGD-100, Perturbation Budget (epsilon_a)=4/255, Patch Granularity=16x162026.05 | 67.9 | 0 | |
| R-TPTBackbone=ViT-B/16, Attack Protocol=PGD-100, Epsilon (ε)=4/255, Evaluation Setting=Zero-shot2026.06 | 67.8 | 47.3 | |
| MTASetting=Zero-shot, Backbone=ViT-B/16, Attack=PGD-100, Perturbation Budget (epsilon_a)=4/255, Patch Granularity=16x162026.05 | 67.5 | 27.5 | |
| CLIP2026.06 | 67.5 | — | |
| R-TPTSetting=Zero-shot, Backbone=ViT-B/16, Attack=PGD-100, Perturbation Budget (epsilon_a)=4/255, Patch Granularity=16x162026.05 | 67.2 | 43.2 | |
| CLIPBackbone=ViT-B/16, Attack Protocol=PGD-100, Epsilon (ε)=4/255, Evaluation Setting=Zero-shot2026.06 | 67 | 0 | |
| FAPShots=16, Training Epochs=10, Training Attack (PGD)=2-step (1/255), Evaluation Attack (PGD-100)=100-step (1/255)2026.05 | 65.7 | 32.8 | |
| C-TPTSetting=Zero-shot, Backbone=ViT-B/16, Attack=PGD-100, Perturbation Budget (epsilon_a)=4/255, Patch Granularity=16x162026.05 | 65.5 | 0 | |
| DBDBackbone=ViT-B/32, Attack Protocol=PGD-100, Epsilon (ε)=4/255, Evaluation Setting=Zero-shot2026.06 | 65.3 | 92.2 | |
| CLIPSetting=Zero-shot, Backbone=ViT-B/16, Attack=PGD-100, Perturbation Budget (epsilon_a)=4/255, Patch Granularity=16x162026.05 | 65.2 | 0 | |
| TTCBackbone=ViT-B/16, Attack Protocol=PGD-100, Epsilon (ε)=4/255, Evaluation Setting=Zero-shot2026.06 | 64.8 | 4.9 | |
| CLIPBackbone=ViT-B/32, Attack Protocol=PGD-100, Epsilon (ε)=4/255, Evaluation Setting=Zero-shot2026.06 | 63.6 | 0 | |
| R-TPTBackbone=ViT-B/32, Attack Protocol=PGD-100, Epsilon (ε)=4/255, Evaluation Setting=Zero-shot2026.06 | 63.1 | 44.2 | |
| TTCBackbone=ViT-B/32, Attack Protocol=PGD-100, Epsilon (ε)=4/255, Evaluation Setting=Zero-shot2026.06 | 60.8 | 27.6 | |
| DBDBackbone=ResNet50, Attack Protocol=PGD-10, Epsilon (ε)=1/255, Evaluation Setting=Zero-shot2026.06 | 60.6 | 88.9 | |
| CLIPBackbone=ResNet50, Attack Protocol=PGD-10, Epsilon (ε)=1/255, Evaluation Setting=Zero-shot2026.06 | 59.7 | 0 | |
| R-TPTBackbone=ResNet50, Attack Protocol=PGD-10, Epsilon (ε)=1/255, Evaluation Setting=Zero-shot2026.06 | 59.2 | 50.3 | |
| CLIPBackbone=CLIP-ResNet50, Augmented Views=63, Adversarial perturbation budget (epsilon)=1.02026.06 | 59 | 0 | |
| SS-TPTBackbone=CLIP-ResNet50, Augmented Views=63, Adversarial perturbation budget (epsilon)=1.0, mode=test-time2026.06 | 58.3 | 51 | |
| APTBackbone=ViT-B/32, Attack Protocol=PGD-100, Epsilon (ε)=4/255, Evaluation Setting=Few-shot2026.06 | 58.2 | 18.9 | |
| APT+TeCoABackbone=CLIP-ResNet50, Augmented Views=63, Adversarial perturbation budget (epsilon)=1.02026.06 | 51.5 | 51.4 | |
| TeCoABackbone=ResNet50, Attack Protocol=PGD-10, Epsilon (ε)=1/255, Evaluation Setting=Zero-shot2026.06 | 40.4 | 21.8 | |
| TeCoABackbone=CLIP-ResNet50, Augmented Views=63, Adversarial perturbation budget (epsilon)=1.02026.06 | 38.4 | 38.1 | |
| TeCoABackbone=ViT-B/32, Attack Protocol=PGD-100, Epsilon (ε)=4/255, Evaluation Setting=Zero-shot2026.06 | 34 | 6.3 | |
| APTBackbone=CLIP-ResNet50, Augmented Views=63, Adversarial perturbation budget (epsilon)=1.02026.06 | 11.2 | 0.9 | |
| AdvVPShots=16, Training Epochs=10, Training Attack (PGD)=2-step (1/255), Evaluation Attack (PGD-100)=100-step (1/255)2026.05 | 0.97 | 0.93 |