Prompt Inversion Attack Defense on Skytrax
88.4Attack AccuracyNo Defense
Evaluation results
| Method | Links | ||
|---|---|---|---|
| No DefenseBackbone model=LLaMA-2-7B, Split layer (k)=42026.06 | 88.4 | 32.9 | |
| PCABackbone model=LLaMA-2-7B, Split layer (k)=4, PCA k=5122026.06 | 88 | 35.7 | |
| FisherBackbone model=LLaMA-2-7B, Split layer (k)=42026.06 | 87.9 | 3,766 | |
| PCABackbone model=LLaMA-2-7B, Split layer (k)=4, PCA k=2562026.06 | 87.2 | 46.5 | |
| NoPeekBackbone model=LLaMA-2-7B, Split layer (k)=42026.06 | 81.7 | 258.6 | |
| GaussianBackbone model=LLaMA-2-7B, Split layer (k)=4, σ=0.12026.06 | 76.4 | 46 | |
| LaplaceBackbone model=LLaMA-2-7B, Split layer (k)=4, b=0.12026.06 | 64.9 | 75.2 | |
| Privacy AdaptersBackbone model=LLaMA-2-7B, Split layer (k)=4, r=256, λ=0.52026.06 | 50.8 | 90.1 | |
| GaussianBackbone model=LLaMA-2-7B, Split layer (k)=4, σ=0.52026.06 | 17.2 | 2,537 | |
| LaplaceBackbone model=LLaMA-2-7B, Split layer (k)=4, b=0.52026.06 | 10.9 | 7,580 | |
| ShredMIBackbone model=LLaMA-2-7B, Split layer (k)=4, r=2562026.06 | 8.2 | 4,091 | |
| No Defense2026.06 | 0.884 | 32.9 | |
| PCAk=5122026.06 | 0.88 | 35.7 | |
| Fisher2026.06 | 0.879 | 3,766 | |
| PCAk=2562026.06 | 0.872 | 46.5 | |
| NoPeek2026.06 | 0.817 | 258.6 | |
| Gaussiansigma=0.12026.06 | 0.764 | 46 | |
| Laplaceb=0.12026.06 | 0.649 | 75.2 | |
| privacy adaptersr=256, lambda=0.52026.06 | 0.508 | 90.1 | |
| Gaussiansigma=0.52026.06 | 0.172 | 2,537 | |
| Laplaceb=0.52026.06 | 0.109 | 7,580 | |
| ShredMIr=2562026.06 | 0.082 | 4,091 |