Image Classification on CIFAR10 (Clean and PGD Robustness)
94.3Clean AccuracyLPF
Evaluation results
| Method | Links | ||
|---|---|---|---|
| LPFBackbone=CLIP-L/14, Method category=Test-Time Defense2026.01 | 94.3 | 71 | |
| RoLI - Full-FTBackbone=Swin Transformer, Initialization=Robust Linear Initialization, Finetuning Protocol=Full Finetuning2023.12 | 94.18 | 60.85 | |
| TTCBackbone=CLIP-L/14, Method category=Test-Time Defense2026.01 | 94.1 | 18.4 | |
| CLIPBackbone=CLIP-L/14, Method category=Original2026.01 | 93.4 | 1.2 | |
| RanLI - Full-FTBackbone=Swin Transformer, Initialization=Random Initialization, Finetuning Protocol=Full Finetuning2023.12 | 93.02 | 59.12 | |
| TTEBackbone=CLIP-L/14, Method category=Test-Time Defense2026.01 | 92.3 | 47 | |
| RoLI - LoRABackbone=Swin Transformer, Initialization=Robust Linear Initialization, Finetuning Protocol=LoRA2023.12 | 92.23 | 58.62 | |
| HDBackbone=CLIP-L/14, Method category=Test-Time Defense2026.01 | 92.1 | 40.5 | |
| RoLI - BiasBackbone=Swin Transformer, Initialization=Robust Linear Initialization, Finetuning Protocol=Bias-only2023.12 | 91.83 | 58.39 | |
| TWINS-ATBackbone=ResNet-502023.12 | 91.24 | 52.73 | |
| HPT-GPDBackbone=ViT-B/32, Proxy CLIP=ViT-B/16, Attack Protocol=PGD-10, Training time per epoch (s)=14862026.01 | 90.74 | 71.14 | |
| CSRBackbone=CLIP-L/14, Method category=Test-Time Defense2026.01 | 90.7 | 83.6 | |
| RoLI - AdapterBackbone=Swin Transformer, Initialization=Robust Linear Initialization, Finetuning Protocol=Adapter2023.12 | 90.31 | 55.69 | |
| RanLI - BiasBackbone=Swin Transformer, Initialization=Random Initialization, Finetuning Protocol=Bias-only2023.12 | 90.25 | 56.94 | |
| RanLI - VPTBackbone=Swin Transformer, Initialization=Random Initialization, Finetuning Protocol=Visual Prompt Tuning2023.12 | 90.24 | 57.38 | |
| RanLI - AdapterBackbone=Swin Transformer, Initialization=Random Initialization, Finetuning Protocol=Adapter2023.12 | 90.23 | 55.7 | |
| R-TPTBackbone=CLIP-L/14, Method category=Test-Time Defense2026.01 | 90.2 | 82.1 | |
| RanLI - LoRABackbone=Swin Transformer, Initialization=Random Initialization, Finetuning Protocol=LoRA2023.12 | 90.02 | 56.88 | |
| Anti-AdvBackbone=CLIP-L/14, Method category=Test-Time Defense2026.01 | 89.3 | 78.5 | |
| FT-StandardBackbone=ViT-B/32, Proxy CLIP=ViT-B/16, Attack Protocol=PGD-10, Training time per epoch (s)=4062026.01 | 88.67 | 24.21 | |
| TTCBackbone=CLIP-B/32, Category=Test-Time Defense2026.01 | 87.6 | 43.3 | |
| CLIPBackbone=ViT-B/32, Proxy CLIP=ViT-B/16, Attack Protocol=PGD-10, Training time per epoch (s)=02026.01 | 87.43 | 42.18 | |
| AutoLoRaBackbone=ResNet-502023.12 | 86.93 | 57.16 | |
| HDBackbone=CLIP-B/32, Category=Test-Time Defense2026.01 | 86.5 | 4.2 | |
| CLIPBackbone=CLIP-B/32, Category=Original2026.01 | 86.1 | 0.6 | |
| CSRBackbone=CLIP-B/32, Category=Test-Time Defense2026.01 | 86.1 | 58.1 | |
| TTEBackbone=CLIP-B/32, Category=Test-Time Defense2026.01 | 86 | 33.7 | |
| LPFBackbone=CLIP-B/32, Category=Test-Time Defense2026.01 | 84.9 | 27.2 | |
| Anti-AdvBackbone=CLIP-B/32, Category=Test-Time Defense2026.01 | 84.4 | 41.8 | |
| RanLI - LinearBackbone=Swin Transformer, Initialization=Random Initialization, Finetuning Protocol=Linear Probing2023.12 | 84.16 | 37.3 | |
| PMGBackbone=ViT-B/32, Proxy CLIP=ViT-B/16, Attack Protocol=PGD-10, Training time per epoch (s)=18172026.01 | 82.8 | 67.06 | |
| R-TPTBackbone=CLIP-B/32, Category=Test-Time Defense2026.01 | 76.7 | 35.1 | |
| FAREBackbone=CLIP-L/14, Method category=Adversarial Fine-tuning2026.01 | 76.6 | 76.5 | |
| TeCoABackbone=CLIP-L/14, Method category=Adversarial Fine-tuning2026.01 | 73.7 | 73.7 | |
| FT-TeCoABackbone=ViT-B/32, Proxy CLIP=ViT-B/16, Attack Protocol=PGD-10, Training time per epoch (s)=11122026.01 | 66.79 | 40.82 | |
| TeCoABackbone=CLIP-B/32, Category=Adversarial Fine-tuning2026.01 | 50.3 | 50.2 | |
| FAREBackbone=CLIP-B/32, Category=Adversarial Fine-tuning2026.01 | 47.2 | 46.9 |