List your company
ResearchFormerly WizWand

Image Classification on Caltech256 (Robustness Metrics)

88.84Accuracy (Clean)AWT
49.37259.618569.86580.1115Mar 20, 2023Oct 1, 2023Apr 14, 2024Oct 26, 2024May 10, 2025Nov 21, 2025Jun 5, 2026
Updated 4mo ago

Evaluation results

MethodLinks
AWTTrain=false2024.0788.84—————
SuS-X-SDTrain=false2024.0787.45—————
VisDescTrain=false2024.0787.16—————
WaffleCLIPTrain=false2024.0787.04—————
POMPTrain=true2024.0786.76—————
CuPLTrain=false2024.0786.3—————
PromptAlignTrain=true2024.0786.05—————
TPTTrain=true2024.0785.71—————
MaPLeTrain=true2024.0785.58—————
CoCoOpTrain=true2024.0785.39—————
PLOT++Train=true2024.0784.3—————
RoLI - Full-FTBackbone=Swin Transformer, Initialization=Robust Linear Initialization, Finetuning Protocol=Full Finetuning2023.1283.7855.42————
CoOpTrain=true2024.0782.91—————
CLIPTrain=false2024.0782.5—————
TTEmode=Test-time Defense, attack=10-step PGD, epsilon=1/2552025.1182.49——60.09——
TTEDefense Category=Test-time Defense2025.1182.48——23.23——
TTEBackbone=CLIP-ViT-B/322026.0682.48——23.23——
RanLI - LinearBackbone=Swin Transformer, Initialization=Random Initialization, Finetuning Protocol=Linear Probing2023.1281.7954.47————
RoLI - LoRABackbone=Swin Transformer, Initialization=Robust Linear Initialization, Finetuning Protocol=LoRA2023.1281.7954.5————
RoLI - BiasBackbone=Swin Transformer, Initialization=Robust Linear Initialization, Finetuning Protocol=Bias-only2023.1281.7854.49————
RoLI - AdapterBackbone=Swin Transformer, Initialization=Robust Linear Initialization, Finetuning Protocol=Adapter2023.1281.7454.32————
CLIPFine-tuning dataset=TinyImageNet, Adversarial training=2-step PGD, Epsilon=1/255, Evaluation protocol=Zero-shot2025.1281.73————13.56
CLIPTraining Data=ImageNet-1k, Adversarial Training=2-step PGD, Epsilon=1/2552025.1281.730.26———7.19
CLIP2025.1281.737.19——8.170.26
CLIPDefense Category=None2025.1181.72——0.12——
CLIPattack=10-step PGD, epsilon=1/2552025.1181.72——8.47——
CLIPBackbone=CLIP-ViT-B/322026.0681.72——0.12——
MACBackbone=CLIP-ViT-B/322026.0681.57——52.56——
RNBackbone=CLIP-ViT-B/322026.0681.25——0.16——
ATACmode=Test-time Defense, attack=10-step PGD, epsilon=1/2552025.1180.72——68.02——
FARETraining Data=ImageNet-1k, Adversarial Training=2-step PGD, Epsilon=1/2552025.1280.5758.87———59.95
UCATTraining Data=ImageNet-1k, Adversarial Training=2-step PGD, Epsilon=1/2552025.1280.2761.64———62.71
RanLI - AdapterBackbone=Swin Transformer, Initialization=Random Initialization, Finetuning Protocol=Adapter2023.1280.2452.73————
FT-StandardBackbone=ViT-B/32, Proxy CLIP=ViT-B/16, Attack Protocol=PGD-10, Training time per epoch (s)=4062026.0179.88——23.82——
RanLI - VPTBackbone=Swin Transformer, Initialization=Random Initialization, Finetuning Protocol=Visual Prompt Tuning2023.1279.7453.07————
TTCmode=Test-time Defense, attack=10-step PGD, epsilon=1/2552025.1179.66——60.11——
HPT-GPDBackbone=ViT-B/32, Proxy CLIP=ViT-B/16, Attack Protocol=PGD-10, Training time per epoch (s)=14862026.0179.63——50.06——
RanLI - BiasBackbone=Swin Transformer, Initialization=Random Initialization, Finetuning Protocol=Bias-only2023.1279.552.43————
CLIPBackbone=ViT-B/32, Proxy CLIP=ViT-B/16, Attack Protocol=PGD-10, Training time per epoch (s)=02026.0179.43——23.76——
Anti-advBackbone=CLIP-ViT-B/322026.0679.4——1.44——
HDBackbone=CLIP-ViT-B/322026.0679.12——0.34——
ATACDefense Category=Test-time Defense2025.1179.1——90.86——
CLIP-FTDefense Category=Adversarial Finetuning2025.1178.53——1.41——
CLIP-FTmode=Adversarial Finetuning, attack=10-step PGD, epsilon=1/2552025.1178.53——6.76——
CLIP-FTBackbone=CLIP-ViT-B/322026.0678.53——1.41——
R-TPTDefense Category=Test-time Defense2025.1177.67——54.45——
R-TPTmode=Test-time Defense, attack=10-step PGD, epsilon=1/2552025.1177.67——67.51——
TWINS-ATwarmup=true2023.0377.3549.1343.92———
RanLI - Full-FTBackbone=Swin Transformer, Initialization=Random Initialization, Finetuning Protocol=Full Finetuning2023.1277.3549.13————
RanLI - LoRABackbone=Swin Transformer, Initialization=Random Initialization, Finetuning Protocol=LoRA2023.1277.0250.25————
TWINS-ATwarmup=false2023.0376.8648.443.69———
TWINS-ATBackbone=ResNet-502023.1276.8648.4————
PMGBackbone=ViT-B/32, Proxy CLIP=ViT-B/16, Attack Protocol=PGD-10, Training time per epoch (s)=18172026.0176.82——41.21——
TTCDefense Category=Test-time Defense2025.1176.59——27.25——
TTCBackbone=CLIP-ViT-B/322026.0676.59——27.25——
TeCoATraining Data=ImageNet-1k, Adversarial Training=2-step PGD, Epsilon=1/2552025.1276.5158.59———59.54
AT2023.0375.948.3743.85———
PMG-AFTFine-tuning dataset=TinyImageNet, Adversarial training=2-step PGD, Epsilon=1/2552025.1273.5943.46——35.6828.29
TWINS-TRADESwarmup=true2023.0373.3948.5343.55———
FAREmode=Adversarial Finetuning, attack=10-step PGD, epsilon=1/2552025.1173.32——38.79——
FAREBackbone=CLIP-ViT-B/32, Adversarial Fine-tuning Attack Budget (ε)=12026.0673.32——2.18——
FT-TeCoABackbone=ViT-B/32, Proxy CLIP=ViT-B/16, Attack Protocol=PGD-10, Training time per epoch (s)=11122026.0172.98——40.56——
FAREFine-tuning dataset=TinyImageNet, Adversarial training=2-step PGD, Epsilon=1/2552025.1272.0520.7——20.855.59
UCATFine-tuning dataset=TinyImageNet, Adversarial training=2-step PGD, Epsilon=1/255, Evaluation protocol=Zero-shot2025.1271.53————48.66
UCATFine-tuning dataset=TinyImageNet, Adversarial training=2-step PGD, Epsilon=1/2552025.1271.5350.47——49.5448.66
TWINS-TRADESwarmup=false2023.0371.1247.3141.77———
AutoLoRaBackbone=ResNet-502023.1269.8547.82————
TRADES2023.0369.747.2843.39———
FAREDefense Category=Adversarial Finetuning2025.1167.22——5.09——
FAREBackbone=CLIP-ViT-B/32, Adversarial Fine-tuning Attack Budget (ε)=42026.0667.22——5.09——
TGA-ZSRFine-tuning dataset=TinyImageNet, Adversarial training=2-step PGD, Epsilon=1/2552025.1266.3146.18——45.8445.16
PMG-AFTmode=Adversarial Finetuning, attack=10-step PGD, epsilon=1/2552025.1162.24——45.91——
PMG-AFTBackbone=CLIP-ViT-B/32, Adversarial Fine-tuning Attack Budget (ε)=12026.0662.24——10.65——
TeCoAmode=Adversarial Finetuning, attack=10-step PGD, epsilon=1/2552025.1161.14——43.19——
TeCoABackbone=CLIP-ViT-B/32, Adversarial Fine-tuning Attack Budget (ε)=12026.0661.14——8.29——
TeCoAFine-tuning dataset=TinyImageNet, Adversarial training=2-step PGD, Epsilon=1/2552025.125942.38——41.741.19
PMG-AFTDefense Category=Adversarial Finetuning2025.1153.32——13.68——
PMG-AFTBackbone=CLIP-ViT-B/32, Adversarial Fine-tuning Attack Budget (ε)=42026.0653.32——13.68——
TeCoADefense Category=Adversarial Finetuning2025.1152.05——11.76——
TeCoABackbone=CLIP-ViT-B/32, Adversarial Fine-tuning Attack Budget (ε)=42026.0652.05——11.76——
DKLFine-tuning dataset=TinyImageNet, Adversarial training=2-step PGD, Epsilon=1/255, Evaluation protocol=Zero-shot2025.1252————38.33
TRADESFine-tuning dataset=TinyImageNet, Adversarial training=2-step PGD, Epsilon=1/255, Evaluation protocol=Zero-shot2025.1250.89————37.84
ACATFine-tuning dataset=TinyImageNet, Adversarial training=2-step PGD, Epsilon=1/255, Evaluation protocol=Zero-shot2025.1250.89————34.84
ATACAttack Strategy=PGD, Epsilon=4/2552025.11———90.86——
ATACAttack Strategy=Lure, Epsilon=4/255, Running time (s)=0.6562025.11———15.83——
ATACAttack Strategy=Avoid, Epsilon=4/255, Running time (s)=0.6582025.11———47.5——
CLIPEvaluation Protocol=Zero-shot, Attack Strategy=AutoAttack, Perturbation Budget=1/2552026.01———0.15——
HPT-GPDEvaluation Protocol=Zero-shot, Attack Strategy=AutoAttack, Perturbation Budget=1/2552026.01———19.52——
PMGEvaluation Protocol=Zero-shot, Attack Strategy=AutoAttack, Perturbation Budget=1/2552026.01———1.54——
TTCAttack Strategy=PGD, Epsilon=4/2552025.11———27.25——
TTCAttack Strategy=Lure, Epsilon=4/255, Running time (s)=0.1492025.11———4.58——
TTCAttack Strategy=Avoid, Epsilon=4/255, Running time (s)=0.1112025.11———8.75——