List your company
ResearchFormerly WizWand

Membership Inference Attack on Asclepius (fine-tuned)

8.17TPR@FPR=0.01HT-MIA
0.60922.57214.5356.4979Jan 27, 2026
Updated 8mo ago

Evaluation results

MethodLinks
HT-MIATarget Model=LLaMA-3.2-1B2026.018.1773.4837.07
HT-MIATarget Model=LLaMA-3.2-1B2026.018.17——
HT-MIATarget Model=LLaMA-3.2-1B, Fine-tuned=true2026.018.1773.4837.07
HT-MIATarget Model=Qwen-3-0.6B2026.014.7773.1231.29
HT-MIATarget Model=Qwen-3-0.6B2026.014.77——
HT-MIATarget Model=Qwen-3-0.6B, Fine-tuned=true2026.014.7773.1231.29
SPV-MIATarget Model=Qwen-3-0.6B2026.014.1——
SPV-MIATarget Model=Qwen-3-0.6B, Fine-tuned=true2026.014.164.9127.2
RatioTarget Model=LLaMA-3.2-1B2026.012.7167.4923.3
RatioTarget Model=LLaMA-3.2-1B2026.012.71——
RatioTarget Model=LLaMA-3.2-1B, Fine-tuned=true2026.012.7167.4923.3
RatioTarget Model=Qwen-3-0.6B2026.012.5——
RatioTarget Model=Qwen-3-0.6B, Fine-tuned=true2026.012.568.1922.05
LowercaseTarget Model=LLaMA-3.2-1B2026.012.4761.3717.96
LowercaseTarget Model=LLaMA-3.2-1B2026.012.47——
LowercaseTarget Model=LLaMA-3.2-1B, Fine-tuned=true2026.012.4761.3717.96
ZlibTarget Model=LLaMA-3.2-1B2026.012.3963.819.77
ZlibTarget Model=LLaMA-3.2-1B2026.012.39——
ZlibTarget Model=LLaMA-3.2-1B, Fine-tuned=true2026.012.3963.819.77
LowercaseTarget Model=Qwen-3-0.6B2026.012.36——
LowercaseTarget Model=Qwen-3-0.6B, Fine-tuned=true2026.012.3661.6617.99
LossTarget Model=LLaMA-3.2-1B2026.012.2863.3219.7
LossTarget Model=LLaMA-3.2-1B2026.012.28——
LossTarget Model=LLaMA-3.2-1B, Fine-tuned=true2026.012.2863.3219.7
ZlibTarget Model=Qwen-3-0.6B2026.012.25——
ZlibTarget Model=Qwen-3-0.6B, Fine-tuned=true2026.012.2563.6519.05
LossTarget Model=Qwen-3-0.6B2026.012.2——
LossTarget Model=Qwen-3-0.6B, Fine-tuned=true2026.012.263.2218.88
SPV-MIATarget Model=LLaMA-3.2-1B2026.012.163.9921.7
SPV-MIATarget Model=LLaMA-3.2-1B2026.012.1——
SPV-MIATarget Model=LLaMA-3.2-1B, Fine-tuned=true2026.012.163.9921.7
PACTarget Model=Qwen-3-0.6B2026.011.94——
PACTarget Model=Qwen-3-0.6B, Fine-tuned=true2026.011.9462.2517.75
PACTarget Model=LLaMA-3.2-1B2026.011.8961.5617.34
PACTarget Model=LLaMA-3.2-1B2026.011.89——
PACTarget Model=LLaMA-3.2-1B, Fine-tuned=true2026.011.8961.5617.34
Min-K++Target Model=LLaMA-3.2-1B2026.011.8260.8416.23
Min-K++Target Model=LLaMA-3.2-1B2026.011.82——
Min-K++Target Model=LLaMA-3.2-1B, Fine-tuned=true2026.011.8260.8416.23
HT-MIATarget Model=GPT-22026.011.8156.7911.09
HT-MIATarget Model=GPT-22026.011.81——
HT-MIATarget Model=GPT-2, Fine-tuned=true2026.011.8156.7911.09
Min-K++Target Model=Qwen-3-0.6B2026.011.49——
Min-K++Target Model=Qwen-3-0.6B, Fine-tuned=true2026.011.4961.0315.45
LowercaseTarget Model=GPT-22026.011.44——
LowercaseTarget Model=GPT-2, Fine-tuned=true2026.011.4454.3112.54
LossTarget Model=GPT-22026.011.31——
ZlibTarget Model=GPT-22026.011.31——
LossTarget Model=GPT-2, Fine-tuned=true2026.011.3153.4712.02
ZlibTarget Model=GPT-2, Fine-tuned=true2026.011.3153.6612.07
Min-K++Target Model=GPT-22026.011.27——
Min-K++Target Model=GPT-2, Fine-tuned=true2026.011.2753.2411.51
RatioTarget Model=GPT-22026.011.2——
RatioTarget Model=GPT-2, Fine-tuned=true2026.011.253.3211.7
PACTarget Model=GPT-22026.011.16——
PACTarget Model=GPT-2, Fine-tuned=true2026.011.165311.57
SPV-MIATarget Model=GPT-22026.010.9——
SPV-MIATarget Model=GPT-2, Fine-tuned=true2026.010.954.4513.1