Input Reconstruction on AlpacaEval (random subset of 100 samples)
99.69Avg Token-level Precision/RecallSipIt
Evaluation results
| Method | Links | ||
|---|---|---|---|
| SipItDefense=No Defense, Target Model=Qwen3-8B, Hardware=RTX 40902026.05 | 99.69 | 4.92 | |
| ActInvDefense=No Defense, Target Model=Qwen3-8B, Hardware=RTX 4090, Note=Experiment Part 22026.05 | 99.65 | 0.78 | |
| ActInvDefense=No Defense, Target Model=Qwen3-8B, Hardware=RTX 4090, Note=Experiment Part 12026.05 | 99.52 | 7.2 | |
| PIADefense=No Defense, Target Model=Qwen3-8B, Hardware=RTX 40902026.05 | 98.48 | 9.59 | |
| ActInvDefense=Sparsification, Target Model=Qwen3-8B, Hardware=RTX 40902026.05 | 96.61 | 7.2 | |
| PIADefense=Sparsification, Target Model=Qwen3-8B, Hardware=RTX 40902026.05 | 95.57 | 9.59 | |
| ActInvDefense=Gaussian Noise, Target Model=Qwen3-8B, Hardware=RTX 40902026.05 | 90.65 | 0.78 | |
| SipItDefense=Gaussian Noise, Target Model=Qwen3-8B, Hardware=RTX 40902026.05 | 90.63 | 4.92 | |
| TBSDefense=No Defense, Target Model=Qwen3-8B, Hardware=RTX 40902026.05 | 88.32 | 7.43 | |
| TBSDefense=Sparsification, Target Model=Qwen3-8B, Hardware=RTX 40902026.05 | 81.97 | 7.43 | |
| A1Defense=No Defense, Target Model=Qwen3-8B, Hardware=RTX 40902026.05 | 43.08 | 0.22 | |
| A1Defense=Sparsification, Target Model=Qwen3-8B, Hardware=RTX 40902026.05 | 18.57 | 0.22 |