Text-to-Text Membership Inference on Wiki103
90ASROurs
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| OursBackbone Model=GPT-2 (1.5B), Knowledge Setting=Partial Knowledge2026.07 | 90 | 0.93 | 52 | |
| ICP-MIABackbone Model=GPT-2 (1.5B), Knowledge Setting=Partial Knowledge2026.07 | 89 | 0.91 | 51 | |
| SPV-MIABackbone Model=GPT-2 (1.5B), Knowledge Setting=Partial Knowledge2026.07 | 87 | 0.88 | 48 | |
| unified membership inference frameworkKnowledge Setting=Zero-Knowledge, Base Model=GPT-2 (1.5B), Embedding Extractor=DistilBERT2026.07 | 83 | 0.9 | 48 | |
| ICPKnowledge Setting=Zero-Knowledge, Base Model=GPT-2 (1.5B), Embedding Extractor=DistilBERT2026.07 | 73 | 0.8 | 39 | |
| SPVKnowledge Setting=Zero-Knowledge, Base Model=GPT-2 (1.5B), Embedding Extractor=DistilBERT2026.07 | 70 | 0.79 | 40 |