Membership Inference on Mistral-7B membership dataset
84.6AUCAttention Flow
Evaluation results
| Method | Links | |
|---|---|---|
| Attention FlowFeatures=per-layer (32-48 dims), Protocol=5-fold CV, Sample size=250 samples2026.05 | 84.6 | |
| CRM-LTSFeatures=per-layer (32-48 dims), Protocol=5-fold CV, Sample size=250 samples2026.05 | 82.5 | |
| Logit LensFeatures=per-layer (32-48 dims), Protocol=5-fold CV, Sample size=250 samples2026.05 | 79.2 | |
| Gradient NormFeatures=per-layer (32-48 dims), Protocol=5-fold CV, Sample size=250 samples2026.05 | 50 |