Dataset Inference on ImageNet 1k
4Required Sample CountOur MIAs for IARs
Evaluation results
| Method | Links | ||
|---|---|---|---|
| Our MIAs for IARsModel variant=VAR-d302025.02 | 4 | -4 | |
| Our MIAs for IARsModel variant=VAR-d242025.02 | 7 | -33 | |
| Optimized ProcedureModel variant=VAR-d302025.02 | 8 | -12 | |
| Our MIAs for IARsModel variant=RAR-XXL2025.02 | 8 | -2 | |
| Optimized ProcedureModel variant=RAR-XXL2025.02 | 10 | -20 | |
| Our MIAs for IARsModel variant=VAR-d202025.02 | 20 | -180 | |
| Baseline (LLM DI)Model variant=VAR-d302025.02 | 20 | — | |
| Our MIAs for IARsModel variant=RAR-XL2025.02 | 20 | -10 | |
| Our MIAs for IARsModel variant=RAR-L2025.02 | 30 | -50 | |
| Optimized ProcedureModel variant=RAR-XL2025.02 | 30 | -10 | |
| Baseline (LLM DI)Model variant=RAR-XXL2025.02 | 30 | — | |
| Optimized ProcedureModel variant=VAR-d242025.02 | 40 | -20 | |
| Baseline (LLM DI)Model variant=RAR-XL2025.02 | 40 | — | |
| Baseline (LLM DI)Model variant=VAR-d242025.02 | 60 | — | |
| Our MIAs for IARsModel variant=RAR-B2025.02 | 80 | -220 | |
| Optimized ProcedureModel variant=RAR-L2025.02 | 80 | -120 | |
| Our MIAs for IARsModel variant=VAR-d162025.02 | 100 | -500 | |
| Optimized ProcedureModel variant=VAR-d202025.02 | 200 | -100 | |
| Baseline (LLM DI)Model variant=RAR-L2025.02 | 200 | — | |
| Baseline (LLM DI)Model variant=VAR-d202025.02 | 300 | — | |
| Our MIAs for IARsModel variant=MAR-H2025.02 | 300 | -500 | |
| Optimized ProcedureModel variant=RAR-B2025.02 | 300 | -200 | |
| Baseline (LLM DI)Model variant=RAR-B2025.02 | 500 | — | |
| Optimized ProcedureModel variant=VAR-d162025.02 | 600 | -1,400 | |
| Our MIAs for IARsModel variant=MAR-L2025.02 | 600 | -1,400 | |
| Optimized ProcedureModel variant=MAR-H2025.02 | 800 | -100 | |
| Baseline (LLM DI)Model variant=MAR-H2025.02 | 900 | — | |
| Baseline (LLM DI)Model variant=VAR-d162025.02 | 2,000 | — | |
| Our MIAs for IARsModel variant=MAR-B2025.02 | 2,000 | -2,000 | |
| Baseline (LLM DI)Model variant=MAR-L2025.02 | 2,000 | — | |
| Optimized ProcedureModel variant=MAR-L2025.02 | 2,000 | 0 | |
| Optimized ProcedureModel variant=MAR-B2025.02 | 4,000 | -1,000 | |
| Baseline (LLM DI)Model variant=MAR-B2025.02 | 5,000 | — |