Training Data Exposure Detection on MedTrinity
80.1AUCGradAudit
Evaluation results
| Method | Links | ||
|---|---|---|---|
| GradAuditModel=Qwen2-VL, Architecture=Generation, Params=2.3B, Training Stage=Fine-tuning2026.06 | 80.1 | 14.5 | |
| GradNormModel=Qwen2-VL, Architecture=Generation, Params=2.3B, Training Stage=Fine-tuning2026.06 | 77.8 | 11.2 | |
| EntropyModel=Qwen2-VL, Architecture=Generation, Params=2.3B, Training Stage=Fine-tuning2026.06 | 58.8 | 10.7 | |
| DC-PDDModel=Qwen2-VL, Architecture=Generation, Params=2.3B, Training Stage=Fine-tuning2026.06 | 55.8 | 6.4 | |
| Min-KModel=Qwen2-VL, Architecture=Generation, Params=2.3B, Training Stage=Fine-tuning2026.06 | 55 | 6.4 | |
| ZlibModel=Qwen2-VL, Architecture=Generation, Params=2.3B, Training Stage=Fine-tuning2026.06 | 54.3 | 5.4 | |
| ModRenyiModel=Qwen2-VL, Architecture=Generation, Params=2.3B, Training Stage=Fine-tuning2026.06 | 54.1 | 6.7 | |
| LossModel=Qwen2-VL, Architecture=Generation, Params=2.3B, Training Stage=Fine-tuning2026.06 | 52.8 | 5.4 | |
| SimilarityModel=Qwen2-VL, Architecture=Generation, Params=2.3B, Training Stage=Fine-tuning2026.06 | 51.8 | 7.41 | |
| M4IModel=Qwen2-VL, Architecture=Generation, Params=2.3B, Training Stage=Fine-tuning2026.06 | 51.5 | 3.6 | |
| Min-K++Model=Qwen2-VL, Architecture=Generation, Params=2.3B, Training Stage=Fine-tuning2026.06 | 50.9 | 4.2 |