Cross-Attack Detection (SSA-CWA vs M-Attack) on LLaVA
99PrecisionDense (Ensemble)
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| Dense (Ensemble)Transfer Setting=SSA-CWA→M-Attack2026.05 | 99 | 100 | 99.5 | |
| SAEgis (vision-block0)Transfer Setting=SSA-CWA→M-Attack2026.05 | 98 | 98 | 98 | |
| SAEgis (Ensemble)Transfer Setting=SSA-CWA→M-Attack2026.05 | 98 | 99 | 98.5 | |
| PIPTransfer Setting=SSA-CWA→M-Attack2026.05 | 97.8 | 93 | 95.3 | |
| SAEgis (projection-mlp2)Transfer Setting=SSA-CWA→M-Attack2026.05 | 96.6 | 58 | 72.5 | |
| DenseTransfer Setting=SSA-CWA→M-Attack2026.05 | 89.7 | 35 | 50.3 |