Image Classification on SVHN (Accuracy and Simulation Metrics)
97AccuracyRL based AL
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| RL based ALModel=VGG-16, Candidate Model Used?=Yes2026.05 | 97 | 0 | — | |
| Linear Feature DisentanglementModel=ResNet-56, Candidate Model Used?=No2026.05 | 96.66 | — | — | |
| Low Confidence (LC)Model=ResNet-56, Candidate Model Used?=No2026.05 | 96.12 | 0.35 | 3.9 | |
| High Confidence and Least Confidence (HCLC)Model=ResNet-56, Candidate Model Used?=No2026.05 | 96.12 | 0.35 | 3.9 | |
| Co-Auxiliary LearningModel=VGG-16, Candidate Model Used?=Yes2026.05 | 96 | 0 | — | |
| High Confidence (HC)Model=ResNet-56, Candidate Model Used?=No2026.05 | 95.99 | 0.35 | 3.9 | |
| Low Confidence (LC)Model=ResNet-18, Candidate Model Used?=No2026.05 | 95.84 | 0.29 | 2.84 | |
| High Confidence and Least Confidence (HCLC)Model=ResNet-18, Candidate Model Used?=No2026.05 | 95.83 | 0.29 | 2.84 | |
| BADGEModel=DenseNet-121, Candidate Model Used?=Yes2026.05 | 95.8 | 0 | 25.92 | |
| Low Confidence (LC)Model=DenseNet-121, Candidate Model Used?=No2026.05 | 95.77 | 0.94 | 16.96 | |
| High Confidence and Least Confidence (HCLC)Model=DenseNet-121, Candidate Model Used?=No2026.05 | 95.76 | 0.94 | 16.96 | |
| High Confidence (HC)Model=ResNet-18, Candidate Model Used?=No2026.05 | 95.65 | 0.29 | 2.84 | |
| High Confidence and Least Confidence (HCLC)Model=VGG-16, Candidate Model Used?=No2026.05 | 95.61 | 0.27 | 4.39 | |
| Low Confidence (LC)Model=VGG-16, Candidate Model Used?=No2026.05 | 95.51 | 0.27 | 4.39 | |
| High Confidence (HC)Model=DenseNet-121, Candidate Model Used?=No2026.05 | 95.48 | 0.94 | 16.96 | |
| High Confidence (HC)Model=VGG-16, Candidate Model Used?=No2026.05 | 95.45 | 0.27 | 4.39 | |
| Beck et al.Model=ResNet-18, Candidate Model Used?=Yes2026.05 | 95.1 | 0 | — | |
| Low Confidence (LC)Model=VGG-16, Candidate Model Used?=No, Budget=50K2026.05 | 94.22 | 0.27 | 3.72 | |
| CUCBModel=DenseNet-121, Candidate Model Used?=Yes2026.05 | 94 | 0 | — | |
| Pretext-based ALModel=VGG-16, Candidate Model Used?=Yes2026.05 | 94 | 0 | — | |
| Low Confidence (LC)Model=ResNet-18, Candidate Model Used?=No, Budget=50K2026.05 | 93.23 | 0.29 | 2.06 | |
| BAL at ScaleModel=VGG-16, Candidate Model Used?=Yes2026.05 | 93 | 0 | — | |
| CoresetModel=ResNet-18, Candidate Model Used?=Yes, Budget=10K2026.05 | 93 | 0 | — | |
| Learning LossModel=ResNet-18, Candidate Model Used?=Yes, Budget=10K2026.05 | 92.5 | 0 | — | |
| CoreGCNModel=ResNet-18, Candidate Model Used?=Yes, Budget=10K2026.05 | 92.5 | 0 | — | |
| UncertainGCNModel=ResNet-18, Candidate Model Used?=Yes, Budget=10K2026.05 | 92.2 | 0 | — | |
| BADGEModel=VGG-16, Candidate Model Used?=Yes, Budget=50K2026.05 | 92 | 0 | — | |
| VAALModel=ResNet-18, Candidate Model Used?=Yes, Budget=10K2026.05 | 92 | 0 | — | |
| Low Confidence (LC)Model=ResNet-18, Candidate Model Used?=No, Budget=15K2026.05 | 91.8 | 0.29 | 0.46 | |
| BAITModel=DenseNet-121, Candidate Model Used?=Yes2026.05 | 91.37 | 0 | 30.06 | |
| ActiveLossNetModel=ResNet-18, Candidate Model Used?=Yes, Budget=10K2026.05 | 91.3 | 0 | — | |
| GlisterModel=DenseNet-121, Candidate Model Used?=Yes2026.05 | 91.26 | 0 | 15.5 | |
| BADGEModel=ResNet-18, Candidate Model Used?=Yes, Budget=50K2026.05 | 91 | 0 | — | |
| Random (Baseline)Model=ResNet-18, Candidate Model Used?=No, Budget=10K2026.05 | 90.5 | 0 | 0.19 | |
| FeatPropModel=ResNet-18, Candidate Model Used?=Yes, Budget=10K2026.05 | 90.5 | 0 | — | |
| ALFA-Mix (large budget)Model=DenseNet-121, Candidate Model Used?=Yes2026.05 | 90 | 0 | — | |
| No ProtectionProtect=-, Attack=-, Backbone=ViT-B/322026.06 | 83.25 | — | — | |
| MergeLockProtect=MergeLock [36], Attack=AGA, Backbone=ViT-B/322026.06 | 81.15 | — | — | |
| BAITModel=ResNet-18, Candidate Model Used?=Yes, Budget=15K2026.05 | 75 | 0 | — | |
| AGAProtect=Params-D, Attack=AGA, Backbone=ViT-B/32, Merging Paradigm=Task Arithmetic2026.06 | 67.36 | — | — | |
| AGAProtect=MergeBarrier, Attack=AGA, Backbone=ViT-B/32, Merging Paradigm=Task Arithmetic2026.06 | 67.34 | — | — | |
| Unprotected BaselineProtect=None, Attack=None, Backbone=ViT-B/32, Merging Paradigm=Task Arithmetic2026.06 | 67.23 | — | — | |
| Double-Cross AttacksModel=ResNet-56, Candidate Model Used?=Yes2026.05 | 67.19 | 0 | — | |
| AGAProtect=Params, Attack=AGA, Backbone=ViT-B/32, Merging Paradigm=Task Arithmetic2026.06 | 67.12 | — | — | |
| AGAProtect=MergeLock, Attack=AGA, Backbone=ViT-B/32, Merging Paradigm=Task Arithmetic2026.06 | 67.11 | — | — | |
| MergeGuardProtect=MergeGuard [5], Attack=AGA, Backbone=ViT-B/322026.06 | 57.12 | — | — | |
| MergeGuardProtect=MergeGuard [5], Attack=-, Backbone=ViT-B/322026.06 | 40.4 | — | — | |
| ARFProtect=ARF, Attack=AGA, Backbone=ViT-B/322026.06 | 37.73 | — | — | |
| MergeBarrierProtect=MergeBarrier, Attack=None, Backbone=ViT-B/32, Merging Paradigm=Task Arithmetic2026.06 | 34.6 | — | — | |
| MergeLockProtect=MergeLock [36], Attack=-, Backbone=ViT-B/322026.06 | 9.05 | — | — | |
| ARFProtect=ARF, Attack=-, Backbone=ViT-B/322026.06 | 8.2 | — | — | |
| Params-DProtect=Params-D, Attack=None, Backbone=ViT-B/32, Merging Paradigm=Task Arithmetic2026.06 | 5.46 | — | — | |
| ParamsProtect=Params, Attack=None, Backbone=ViT-B/32, Merging Paradigm=Task Arithmetic2026.06 | 5.05 | — | — | |
| MergeLockProtect=MergeLock, Attack=None, Backbone=ViT-B/32, Merging Paradigm=Task Arithmetic2026.06 | 4.99 | — | — |