Adversarial Attack Performance Evaluation on API-based two-layer LLM cascade v1 (test)
93.5Cascade (Φ) AccuracyInitial cascade
Evaluation results
| Method | Links | |||||||
|---|---|---|---|---|---|---|---|---|
| Initial cascadeconfiguration=(21)2026.05 | 93.5 | 0.52 | 82.5 | 0.27 | 94.5 | 1.37 | 98.5 | |
| Initial cascadeconfiguration=(23)2026.05 | 91 | 0.19 | 80 | 0.092 | 92 | 0.57 | 91 | |
| Joint-target Attackp=0, Transfer Source=1, Target Cascade=212026.05 | 76 | 0.44 | 75 | 0.29 | 87 | 1.42 | 73 | |
| Joint-target Attackp=0, Transfer Source=1, Target Cascade=232026.05 | 75.5 | 0.14 | 72 | 0.097 | 89 | 0.6 | 71.5 | |
| Joint-target Attackp=0, Transfer Source=8, Target Cascade=212026.05 | 74 | 0.42 | 71 | 0.28 | 89 | 1.41 | 75 |