Membership Inference Attack on SNLI
99.8ROC AUCRMIA
Evaluation results
| Method | Links | |
|---|---|---|
| RMIAThreat Model=Model-based (A^theta), Source=Synthetic, Label=Natural, Canary Length=50 words2025.02 | 99.8 | |
| RMIAThreat Model=Model-based (A^theta), Source=Synthetic, Label=Artificial, Canary Length=50 words2025.02 | 99.7 | |
| RMIAThreat Model=Model-based (A^theta), Source=In-distribution, Canary Length=30 words2025.02 | 89.2 | |
| n-gram MIAThreat Model=Data-based (A^D), Methodology=2-gram, Source=Synthetic, Label=Artificial, Canary Length=50 words2025.02 | 77 | |
| n-gram MIAThreat Model=Data-based (A^D), Methodology=2-gram, Source=In-distribution, Canary Length=30 words2025.02 | 71.8 | |
| similarity-based MIA (SIMJac)Threat Model=Data-based (A^D), Methodology=SIMJac, Source=In-distribution, Canary Length=30 words2025.02 | 64.4 | |
| similarity-based MIA (SIMemb)Threat Model=Data-based (A^D), Methodology=SIMemb, Source=In-distribution, Canary Length=30 words2025.02 | 63 | |
| similarity-based MIA (SIMJac)Threat Model=Data-based (A^D), Methodology=SIMJac, Source=Synthetic, Label=Artificial, Canary Length=50 words2025.02 | 60.2 | |
| similarity-based MIA (SIMemb)Threat Model=Data-based (A^D), Methodology=SIMemb, Source=Synthetic, Label=Artificial, Canary Length=50 words2025.02 | 57.1 | |
| n-gram MIAThreat Model=Data-based (A^D), Methodology=2-gram, Source=Synthetic, Label=Natural, Canary Length=50 words2025.02 | 53.4 | |
| similarity-based MIA (SIMemb)Threat Model=Data-based (A^D), Methodology=SIMemb, Source=Synthetic, Label=Natural, Canary Length=50 words2025.02 | 48.8 | |
| similarity-based MIA (SIMJac)Threat Model=Data-based (A^D), Methodology=SIMJac, Source=Synthetic, Label=Natural, Canary Length=50 words2025.02 | 48.6 |