Image Classification on CIFAR-100 (test) (Accuracy, ECE, AdaECE, CW-ECE)
92.46AccuracyCCL-SC
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| CCL-SCBackbone=ViT, Training Protocol=transfer learning, Epochs=502026.04 | 92.46 | 0.64 | 0.47 | 15 | |
| SocratesBackbone=ViT, Training Protocol=transfer learning, Epochs=502026.04 | 91.83 | 5.3 | 5.3 | 17 | |
| SocratesBackbone=ViT, Training Protocol=Transfer Learning (TL)2026.04 | 91.83 | 5.3 | 5.3 | 17 | |
| Socrates+VSBackbone=ViT, Training Protocol=Transfer Learning (TL)2026.04 | 91.47 | 1.77 | 1.7 | 14 | |
| CEModel=ViT-Small, Pre-training=ImageNet-pretrained, Fine-tuning=Full-parameter2025.08 | 90.77 | 3.25 | — | — | |
| FGRModel=ViT-Small, Pre-training=ImageNet-pretrained, Fine-tuning=Full-parameter2025.08 | 90.4 | 0.93 | — | — | |
| DFLModel=ViT-Small, Pre-training=ImageNet-pretrained, Fine-tuning=Full-parameter2025.08 | 90.24 | 1.39 | — | — | |
| CEModel=Swin-Tiny, Pre-training=ImageNet-pretrained, Fine-tuning=Full-parameter2025.08 | 87.77 | 2.87 | — | — | |
| CEModel=DeiT-Small, Pre-training=ImageNet-pretrained, Fine-tuning=Full-parameter2025.08 | 87.67 | 3.05 | — | — | |
| DFLModel=DeiT-Small, Pre-training=ImageNet-pretrained, Fine-tuning=Full-parameter2025.08 | 87.37 | 4.61 | — | — | |
| FGRModel=DeiT-Small, Pre-training=ImageNet-pretrained, Fine-tuning=Full-parameter2025.08 | 87.29 | 2.21 | — | — | |
| DFLModel=Swin-Tiny, Pre-training=ImageNet-pretrained, Fine-tuning=Full-parameter2025.08 | 87.27 | 0.99 | — | — | |
| FGRModel=Swin-Tiny, Pre-training=ImageNet-pretrained, Fine-tuning=Full-parameter2025.08 | 87.12 | 0.8 | — | — | |
| CCL-SCBackbone=ResNet-110, Training Protocol=standard, Epochs=3002026.04 | 77.85 | 10.4 | 10.4 | 27 | |
| SocratesBackbone=ResNet-110, Training Protocol=standard, Epochs=3002026.04 | 77.39 | 2.86 | 2.76 | 27 | |
| SocratesBackbone=ResNet-110, Training Protocol=Standard2026.04 | 77.39 | 2.86 | 2.76 | 27 | |
| Socrates+VSBackbone=ResNet-110, Training Protocol=Standard2026.04 | 77.31 | 4.44 | 4.4 | 26 | |
| FLSDBackbone=ViT, Training Protocol=transfer learning, Epochs=502026.04 | 77.15 | 1.68 | 1.63 | 29 | |
| FLSDBackbone=ResNet-110, Training Protocol=standard, Epochs=3002026.04 | 77.05 | 4.15 | 3.94 | 27 | |
| FocalBackbone=ResNet-110, Training Protocol=standard, Epochs=3002026.04 | 76.62 | 6.49 | 6.46 | 31 | |
| EnsembleFusion Type=4-way fusion, Evaluation Protocol=Zero-shot, Inference Cost=×42025.06 | 76.6 | — | — | — | |
| CE+VSBackbone=ViT, Training Protocol=Transfer Learning (TL)2026.04 | 75.85 | 1.96 | 1.9 | 26 | |
| EnsembleFusion Type=2-way fusion, Evaluation Protocol=Zero-shot, Inference Cost=×22025.06 | 75.7 | — | — | — | |
| KF-GradientFusion Type=4-way fusion, Evaluation Protocol=Fine-tuning, Inference Cost=×12025.06 | 75.6 | — | — | — | |
| SATBackbone=ResNet-110, Training Protocol=standard, Epochs=3002026.04 | 75.41 | 8.6 | 8.6 | 33 | |
| KF-GradientFusion Type=2-way fusion, Evaluation Protocol=Fine-tuning, Inference Cost=×12025.06 | 75.4 | — | — | — | |
| CSAM2025.05 | 74.98 | 1.25 | — | — | |
| SAM2025.05 | 74.63 | 1.87 | — | — | |
| MCBackbone=ResNet-110, Training Protocol=standard, Epochs=3002026.04 | 74.55 | 13.51 | 13.51 | 30 | |
| BrierBackbone=ResNet-110, Training Protocol=standard, Epochs=3002026.04 | 74.31 | 4.33 | 4.18 | 30 | |
| Base ModelsFusion Type=4-way fusion, Evaluation Protocol=Zero-shot, Inference Cost=×12025.06 | 74.1 | — | — | — | |
| Transformer OTFusionFusion Type=2-way fusion, Evaluation Protocol=Fine-tuning, Inference Cost=×12025.06 | 74 | — | — | — | |
| Base ModelsFusion Type=2-way fusion, Evaluation Protocol=Zero-shot, Inference Cost=×12025.06 | 73.9 | — | — | — | |
| Transformer OTFusionFusion Type=4-way fusion, Evaluation Protocol=Fine-tuning, Inference Cost=×12025.06 | 72.6 | — | — | — | |
| LS2025.05 | 72.54 | 4.38 | — | — | |
| ACLS2025.05 | 72.49 | 2.32 | — | — | |
| CPC2025.05 | 72.45 | 2.62 | — | — | |
| MBLS2025.05 | 72.42 | 2.33 | — | — | |
| CCL-SCBackbone=VGG-16, Training Protocol=standard, Epochs=3002026.04 | 72.41 | 11.91 | 11.87 | 32 | |
| MDCA2025.05 | 72.38 | 3.16 | — | — | |
| CRL2025.05 | 72.36 | 3.43 | — | — | |
| SGD2025.05 | 72.3 | 13.03 | — | — | |
| FocalBackbone=VGG-16, Training Protocol=standard, Epochs=3002026.04 | 71.93 | 7.48 | 7.49 | 28 | |
| Socrates+VSBackbone=VGG-16, Training Protocol=Standard2026.04 | 71.43 | 2.67 | 2.66 | 28 | |
| SATBackbone=ViT, Training Protocol=transfer learning, Epochs=502026.04 | 71.27 | 1.46 | 1.44 | 34 | |
| SocratesBackbone=VGG-16, Training Protocol=standard, Epochs=3002026.04 | 71.26 | 3.45 | 3.51 | 28 | |
| SocratesBackbone=VGG-16, Training Protocol=Standard2026.04 | 71.26 | 3.45 | 3.51 | 28 | |
| CEBackbone=ViT, Training Protocol=Transfer Learning (TL)2026.04 | 70.98 | 1.49 | 1.52 | 34 | |
| FLSDBackbone=VGG-16, Training Protocol=standard, Epochs=3002026.04 | 70.15 | 5.36 | 5.33 | 30 | |
| FocalBackbone=ViT, Training Protocol=transfer learning, Epochs=502026.04 | 69.05 | 6.55 | 6.6 | 39 | |
| MCBackbone=VGG-16, Training Protocol=standard, Epochs=3002026.04 | 68.08 | 8 | 8 | 30 | |
| Safe Residual EstimatorNumber of labeled samples (n)=2000, Backbone=CLIP2026.06 | 66.7 | — | — | — | |
| SATBackbone=VGG-16, Training Protocol=standard, Epochs=3002026.04 | 66.14 | 12.23 | 12.4 | 49 | |
| Safe Residual EstimatorNumber of labeled samples (n)=1000, Backbone=CLIP2026.06 | 64.4 | — | — | — | |
| Weighted Ensemble (Validation-Tuned)Number of labeled samples (n)=2000, Backbone=CLIP2026.06 | 63.9 | — | — | — | |
| KF-GradientFusion Type=2-way fusion, Evaluation Protocol=Zero-shot, Inference Cost=×12025.06 | 63 | — | — | — | |
| MCBackbone=ViT, Training Protocol=transfer learning, Epochs=502026.04 | 62.48 | 1.6 | 1.66 | 43 | |
| Weighted Ensemble (Validation-Tuned)Number of labeled samples (n)=1000, Backbone=CLIP2026.06 | 61.8 | — | — | — | |
| Safe Residual EstimatorNumber of labeled samples (n)=500, Backbone=CLIP2026.06 | 61 | — | — | — | |
| Safe Residual EstimatorNumber of labeled samples (n)=200, Backbone=CLIP2026.06 | 60.4 | — | — | — | |
| Safe Residual EstimatorNumber of labeled samples (n)=100, Backbone=CLIP2026.06 | 59.8 | — | — | — | |
| Weighted Ensemble (Validation-Tuned)Number of labeled samples (n)=500, Backbone=CLIP2026.06 | 59.4 | — | — | — | |
| CLIP Zero-shotBackbone=CLIP, Mode=Zero-shot2026.06 | 59.39 | — | — | — | |
| Weighted Ensemble (Validation-Tuned)Number of labeled samples (n)=100, Backbone=CLIP2026.06 | 59.3 | — | — | — | |
| BrierBackbone=ViT, Training Protocol=transfer learning, Epochs=502026.04 | 58.55 | 7.08 | 7.1 | 52 | |
| KF-GradientFusion Type=4-way fusion, Evaluation Protocol=Zero-shot, Inference Cost=×12025.06 | 57.5 | — | — | — | |
| Scratch (Linear Probe)Number of labeled samples (n)=2000, Backbone=CLIP2026.06 | 57.2 | — | — | — | |
| Weighted Ensemble (Validation-Tuned)Number of labeled samples (n)=200, Backbone=CLIP2026.06 | 54.5 | — | — | — | |
| BrierBackbone=VGG-16, Training Protocol=standard, Epochs=3002026.04 | 53.87 | 6.47 | 7.06 | 51 | |
| FLSDBackbone=ViT, Training Protocol=standard, Epochs=3002026.04 | 49.11 | 10.58 | 10.58 | 49 | |
| SocratesBackbone=ViT, Training Protocol=standard, Epochs=3002026.04 | 49.07 | 4.1 | 4.08 | 45 | |
| SocratesBackbone=ViT, Training Protocol=Standard2026.04 | 49.07 | 4.1 | 4.08 | 45 | |
| FocalBackbone=ViT, Training Protocol=standard, Epochs=3002026.04 | 48.3 | 8.87 | 8.86 | 50 | |
| Socrates+VSBackbone=ViT, Training Protocol=Standard2026.04 | 47.91 | 2.43 | 2.46 | 47 | |
| CCL-SCBackbone=ViT, Training Protocol=standard, Epochs=3002026.04 | 47.63 | 11.66 | 11.64 | 53 | |
| Concat FeaturesNumber of labeled samples (n)=2000, Backbone=CLIP2026.06 | 45.2 | — | — | — | |
| MCBackbone=ViT, Training Protocol=standard, Epochs=3002026.04 | 45.09 | 6.65 | 6.63 | 54 | |
| Scratch (Linear Probe)Number of labeled samples (n)=1000, Backbone=CLIP2026.06 | 42.9 | — | — | — | |
| BrierBackbone=ViT, Training Protocol=standard, Epochs=3002026.04 | 42.85 | 2.09 | 1.96 | 54 | |
| SATBackbone=ViT, Training Protocol=standard, Epochs=3002026.04 | 42.82 | 21.91 | 21.91 | 73 | |
| VagueGANData Distribution=IID, Malicious Clients Percentage α=30%2026.05 | 39.4 | — | — | — | |
| PCDMData Distribution=IID, Malicious Clients Percentage α=30%2026.05 | 37.3 | — | — | — | |
| PCDMData Distribution=non-IID, Malicious Clients Percentage α=30%2026.05 | 37.2 | — | — | — | |
| VagueGANData Distribution=non-IID, Malicious Clients Percentage α=30%2026.05 | 35.1 | — | — | — | |
| Concat FeaturesNumber of labeled samples (n)=1000, Backbone=CLIP2026.06 | 33.3 | — | — | — | |
| Scratch (Linear Probe)Number of labeled samples (n)=500, Backbone=CLIP2026.06 | 21.7 | — | — | — | |
| Concat FeaturesNumber of labeled samples (n)=500, Backbone=CLIP2026.06 | 21.7 | — | — | — | |
| Concat FeaturesNumber of labeled samples (n)=200, Backbone=CLIP2026.06 | 13.2 | — | — | — | |
| Scratch (Linear Probe)Number of labeled samples (n)=200, Backbone=CLIP2026.06 | 9.9 | — | — | — | |
| Transformer OTFusionFusion Type=2-way fusion, Evaluation Protocol=Zero-shot, Inference Cost=×12025.06 | 4.3 | — | — | — | |
| Scratch (Linear Probe)Number of labeled samples (n)=100, Backbone=CLIP2026.06 | 2.1 | — | — | — | |
| Transformer OTFusionFusion Type=4-way fusion, Evaluation Protocol=Zero-shot, Inference Cost=×12025.06 | 1 | — | — | — | |
| Concat FeaturesNumber of labeled samples (n)=100, Backbone=CLIP2026.06 | 0.4 | — | — | — |