Adversarial Robustness on CUB200 (Image Classification)
74.28Accuracy (Clean)RoLI - Full-FT
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| RoLI - Full-FTBackbone=Swin Transformer, Initialization=Robust Linear Initialization, Finetuning Protocol=Full Finetuning2023.12 | 74.28 | 32.03 | — | |
| RanLI - Full-FTBackbone=Swin Transformer, Initialization=Random Initialization, Finetuning Protocol=Full Finetuning2023.12 | 70.83 | 31.55 | — | |
| RanLI - LinearBackbone=Swin Transformer, Initialization=Random Initialization, Finetuning Protocol=Linear Probing2023.12 | 69.88 | 24.56 | — | |
| RoLI - LoRABackbone=Swin Transformer, Initialization=Robust Linear Initialization, Finetuning Protocol=LoRA2023.12 | 68.88 | 26.77 | — | |
| RoLI - BiasBackbone=Swin Transformer, Initialization=Robust Linear Initialization, Finetuning Protocol=Bias-only2023.12 | 68.14 | 25.7 | — | |
| TWINS-ATwarmup=false2023.03 | 68.09 | 29.24 | 22.33 | |
| TWINS-ATBackbone=ResNet-502023.12 | 68.09 | 29.24 | — | |
| TWINS-ATwarmup=true2023.03 | 67.64 | 27.67 | 23.58 | |
| RanLI - BiasBackbone=Swin Transformer, Initialization=Random Initialization, Finetuning Protocol=Bias-only2023.12 | 66.69 | 29.88 | — | |
| AT2023.03 | 65.74 | 26.6 | 22.82 | |
| RanLI - LoRABackbone=Swin Transformer, Initialization=Random Initialization, Finetuning Protocol=LoRA2023.12 | 64.74 | 29.65 | — | |
| RoLI - AdapterBackbone=Swin Transformer, Initialization=Robust Linear Initialization, Finetuning Protocol=Adapter2023.12 | 64.2 | 25.22 | — | |
| AutoLoRaBackbone=ResNet-502023.12 | 62.78 | 31.07 | — | |
| RanLI - VPTBackbone=Swin Transformer, Initialization=Random Initialization, Finetuning Protocol=Visual Prompt Tuning2023.12 | 61.94 | 26.46 | — | |
| TWINS-TRADESwarmup=true2023.03 | 61.05 | 26.68 | 21.92 | |
| TWINS-TRADESwarmup=false2023.03 | 60.72 | 27.05 | 22.68 | |
| RanLI - AdapterBackbone=Swin Transformer, Initialization=Random Initialization, Finetuning Protocol=Adapter2023.12 | 59.67 | 26.73 | — | |
| TRADES2023.03 | 58.92 | 27.87 | 22.21 |