Image Classification on ImageNet-1K (Accuracy, ASR, FR, ACR)
86.7AccuracyDinoV2
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| DinoV2Category=Self-Supervised Pre-training, Arch.=ViT-L, Backbone Status=Frozen, Model capacity=Larger-capacity2025.06 | 86.7 | — | — | — | |
| MAE + LLECategory=Self-Sup. + De-biasing, Arch.=ViT-L, Model capacity=Larger-capacity2025.06 | 85.8 | — | — | — | |
| SWAG + LLECategory=Self-Sup. + De-biasing, Arch.=ViT-B2025.06 | 85.2 | — | — | — | |
| ERMCategory=Supervised Training, Arch.=ViT-L, Model capacity=Larger-capacity2025.06 | 84.8 | — | — | — | |
| DinoV2Category=Self-Supervised Pre-training, Arch.=ViT-B, Backbone Status=Frozen2025.06 | 84.6 | — | — | — | |
| iFAMCategory=Early Masking, Arch.=ViT-B, K=12025.06 | 84.3 | — | — | — | |
| ERMCategory=Supervised Training, Arch.=ViT-B2025.06 | 83.8 | — | — | — | |
| MAE + LLECategory=Self-Sup. + De-biasing, Arch.=ViT-B2025.06 | 83.7 | — | — | — | |
| ERMCategory=Supervised Training, Arch.=R-152, Model capacity=Larger-capacity2025.06 | 83.5 | — | — | — | |
| PDiscoFormerCategory=Late Masking, Arch.=ViT-B, K=12025.06 | 83.3 | — | — | — | |
| Immuno-VLMBackbone=CLIP-ViT-L/14, Protocol=Generative / Training-based2026.05 | 82.1 | — | — | — | |
| BenignVictim Architecture ID=(j), Victim Model Category=CNN2025.06 | 81.49 | — | — | — | |
| ERMCategory=Supervised Training, Arch.=R-502025.06 | 81.2 | — | — | — | |
| BenignVictim Architecture ID=(n), Victim Model Category=Transformer2025.06 | 81.19 | — | — | — | |
| BenignVictim Architecture ID=(l), Victim Model Category=Transformer2025.06 | 80.67 | — | — | — | |
| BenignVictim Architecture ID=(o), Victim Model Category=Transformer2025.06 | 80.48 | — | — | — | |
| BenignVictim Architecture ID=(m), Victim Model Category=Transformer2025.06 | 79.28 | — | — | — | |
| BenignVictim Architecture ID=(i), Victim Model Category=CNN2025.06 | 79.12 | — | — | — | |
| BenignVictim Architecture ID=(p), Victim Model Category=Transformer2025.06 | 79.1 | — | — | — | |
| Immuno-VLMBackbone=CLIP-ViT-B/16, Protocol=Generative / Training-based2026.05 | 78.5 | — | — | — | |
| MSPBackbone=CLIP-ViT-B/16, Protocol=Zero-Shot2026.05 | 78.2 | — | — | — | |
| Energy ScoreBackbone=CLIP-ViT-B/16, Protocol=Zero-Shot2026.05 | 78.2 | — | — | — | |
| MCMBackbone=CLIP-ViT-B/16, Protocol=Zero-Shot2026.05 | 78.2 | — | — | — | |
| ASH-BBackbone=CLIP-ViT-B/16, Protocol=Post-Hoc Adaptation2026.05 | 78 | — | — | — | |
| BenignVictim Architecture ID=(f), Victim Model Category=CNN2025.06 | 77.95 | — | — | — | |
| ReActBackbone=CLIP-ViT-B/16, Protocol=Post-Hoc Adaptation2026.05 | 77.9 | — | — | — | |
| Neg-Label (Standard AIS)Backbone=CLIP-ViT-B/16, Protocol=Generative / Training-based2026.05 | 77.5 | — | — | — | |
| BenignVictim Architecture ID=(b), Victim Model Category=CNN2025.06 | 77.33 | — | — | — | |
| CSIBackbone=CLIP-ViT-B/16, Protocol=Generative / Training-based2026.05 | 76.5 | — | — | — | |
| LLECategory=Specialized De-biasing, Arch.=R-502025.06 | 76.3 | — | — | — | |
| BenignVictim Architecture ID=(e), Victim Model Category=CNN2025.06 | 76.19 | — | — | — | |
| BenignVictim Architecture ID=(d), Victim Model Category=CNN2025.06 | 75.74 | — | — | — | |
| BenignVictim Architecture ID=(k), Victim Model Category=CNN2025.06 | 75.42 | — | — | — | |
| BenignVictim Architecture ID=(a), Victim Model Category=CNN2025.06 | 74.6 | — | — | — | |
| BenignVictim Architecture ID=(c), Victim Model Category=CNN2025.06 | 74.22 | — | — | — | |
| BenignVictim Architecture ID=(v), Victim Model Category=Avg.2025.06 | 73.77 | — | — | — | |
| BenignVictim Architecture ID=(u), Victim Model Category=Mamba2025.06 | 73.21 | — | — | — | |
| BenignVictim Architecture ID=(r), Victim Model Category=Mixer2025.06 | 69.9 | — | — | — | |
| ViT-B/16Architecture=ViT-B/16, Zero-shot=true, Role=Teacher2026.03 | 67.1 | — | — | — | |
| BenignVictim Architecture ID=(s), Victim Model Category=Mixer2025.06 | 66.53 | — | — | — | |
| BenignVictim Architecture ID=(t), Victim Model Category=Mamba2025.06 | 66.53 | — | — | — | |
| BenignVictim Architecture ID=(g), Victim Model Category=CNN2025.06 | 66.5 | — | — | — | |
| BenignVictim Architecture ID=(q), Victim Model Category=Mixer2025.06 | 57.91 | — | — | — | |
| BenignVictim Architecture ID=(h), Victim Model Category=CNN2025.06 | 55.91 | — | — | — | |
| CLIP-RDArchitecture=ResNet-50, Zero-shot=true, Distillation Method=CLIP-RD2026.03 | 55.1 | — | — | — | |
| CLIP-KDArchitecture=ResNet-50, Zero-shot=true, Distillation Method=CLIP-KD2026.03 | 54.1 | — | — | — | |
| CLIP-RDArchitecture=EfficientNet-B0, Zero-shot=true, Distillation Method=CLIP-RD2026.03 | 46.1 | — | — | — | |
| CLIP-RDArchitecture=ResNet-18, Zero-shot=true, Distillation Method=CLIP-RD2026.03 | 38.6 | — | — | — | |
| CLIP-KDArchitecture=ResNet-18, Zero-shot=true, Distillation Method=CLIP-KD2026.03 | 38.2 | — | — | — | |
| ResNet-50Architecture=ResNet-50, Zero-shot=true, Distillation Method=None (Baseline student)2026.03 | 34.8 | — | — | — | |
| EfficientNet-B0Architecture=EfficientNet-B0, Zero-shot=true, Distillation Method=None (Baseline student)2026.03 | 30.9 | — | — | — | |
| ResNet-18Architecture=ResNet-18, Zero-shot=true, Distillation Method=None (Baseline student)2026.03 | 27.3 | — | — | — | |
| BIA + SCGAVictim Architecture ID=(v), Victim Model Category=Avg.2025.06 | — | 1.47 | 1.28 | -0.21 | |
| CDA + SCGAVictim Architecture ID=(v), Victim Model Category=Avg.2025.06 | — | 8.55 | 7.49 | -1.57 | |
| FACL + SCGAVictim Architecture ID=(v), Victim Model Category=Avg.2025.06 | — | 1.35 | 1.2 | -0.28 | |
| GAMA + SCGAVictim Architecture ID=(v), Victim Model Category=Avg.2025.06 | — | 1.44 | 1.28 | -0.14 | |
| LTP + SCGAVictim Architecture ID=(v), Victim Model Category=Avg.2025.06 | — | 8.47 | 7.65 | -0.7 | |
| PDCL + SCGAVictim Architecture ID=(v), Victim Model Category=Avg.2025.06 | — | 0.09 | 0.11 | -0.03 |