Membership Inference Attack on C4 Vocabulary Size: 170,000
0.761AUCVocabulary Overlap
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| Vocabulary OverlapShadow Tokenizers=true, Auxiliary Datasets=true2025.10 | 0.761 | 70.9 | 32.53 | |
| Frequency EstimationShadow Tokenizers=true, Auxiliary Datasets=true2025.10 | 0.707 | 68.1 | 25.61 | |
| Naive BayesShadow Tokenizers=false, Auxiliary Datasets=true, k=60,0002025.10 | 0.553 | 54.5 | 4.94 | |
| Naive BayesShadow Tokenizers=false, Auxiliary Datasets=true, k=40,0002025.10 | 0.55 | 54.2 | 7.6 | |
| Naive BayesShadow Tokenizers=false, Auxiliary Datasets=true, k=20,0002025.10 | 0.546 | 53.8 | 5.86 | |
| Compression RateShadow Tokenizers=false, Auxiliary Datasets=false2025.10 | 0.508 | 51.3 | 0.82 | |
| Merge SimilarityShadow Tokenizers=true, Auxiliary Datasets=true2025.10 | 0.495 | 50.6 | 0.92 |