Membership Inference Attack on Medical-o1 reasoning-SFT
0.9851AUCAll3Layers (MINT-based)
Evaluation Results
| Method | Links | |
|---|---|---|
| All3Layers (MINT-based)Model=DeepSeek-V32026.06 | 0.9851 | |
| L3 (MINT-based)Model=DeepSeek-V32026.06 | 0.9846 | |
| L2 (MINT-based)Model=DeepSeek-V32026.06 | 0.984 | |
| All3Layers (MINT-based)Model=Phi-42026.06 | 0.9819 | |
| L3 (MINT-based)Model=Phi-42026.06 | 0.9811 | |
| L1 (MINT-based)Model=Phi-42026.06 | 0.98 | |
| L2 (MINT-based)Model=Phi-42026.06 | 0.9788 | |
| L1 (MINT-based)Model=DeepSeek-V32026.06 | 0.9697 | |
| MinK++Model=Phi-42026.06 | 0.959 | |
| MinKModel=Phi-42026.06 | 0.944 | |
| GradProbLossModel=DeepSeek-V32026.06 | 0.9435 | |
| LossModel=Phi-42026.06 | 0.939 | |
| All3Layers (MINT-based)Model=Qwen32026.06 | 0.9302 | |
| L3 (MINT-based)Model=Qwen32026.06 | 0.9272 | |
| All3Layers (MINT-based)Model=LLaMA 3.2-3B2026.06 | 0.924 | |
| L3 (MINT-based)Model=LLaMA 3.2-3B2026.06 | 0.921 | |
| L2 (MINT-based)Model=LLaMA 3.2-3B2026.06 | 0.9202 | |
| L2 (MINT-based)Model=Qwen32026.06 | 0.9179 | |
| MinKModel=DeepSeek-V32026.06 | 0.884 | |
| LossModel=DeepSeek-V32026.06 | 0.884 | |
| L1 (MINT-based)Model=LLaMA 3.2-3B2026.06 | 0.8811 | |
| MinK++Model=DeepSeek-V32026.06 | 0.8625 | |
| GradProbLossModel=LLaMA 3.2-3B2026.06 | 0.8394 | |
| L1 (MINT-based)Model=Qwen32026.06 | 0.8306 | |
| LossModel=LLaMA 3.2-3B2026.06 | 0.8 | |
| GradProbLossModel=Qwen32026.06 | 0.792 | |
| MinK++Model=LLaMA 3.2-3B2026.06 | 0.7812 | |
| MinK++Model=Qwen32026.06 | 0.778 | |
| MinKModel=LLaMA 3.2-3B2026.06 | 0.777 | |
| MinKModel=Qwen32026.06 | 0.769 | |
| LossModel=Qwen32026.06 | 0.7508 | |
| OnlyGradModel=Phi-42026.06 | 0.7132 | |
| OnlyGradModel=Qwen32026.06 | 0.6706 | |
| OnlyGradModel=DeepSeek-V32026.06 | 0.6683 | |
| GradProbLossModel=Phi-42026.06 | 0.6323 | |
| OnlyGradModel=LLaMA 3.2-3B2026.06 | 0.5927 |