Trigger Search on Backdoored Models Token-Level BERT-based (test)
100Recall (NeuBA)Patronus
Evaluation Results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| PatronusFuzzy Search Rounds (FS)=5, Optimization Strategy=beam search2025.12 | 100 | 12.18 | 100 | 31.1 | 98.85 | 31.79 | |
| PatronusFuzzy Search Rounds (FS)=3, Optimization Strategy=beam search2025.12 | 94.05 | 7.25 | 95.32 | 18.96 | 96.55 | 18.6 | |
| Patronus-greedyFuzzy Search Rounds (FS)=5, Optimization Strategy=greedy optimization2025.12 | 92.86 | 1.85 | 89.47 | 4.39 | 93.68 | 4.39 | |
| PatronusFuzzy Search Rounds (FS)=2, Optimization Strategy=beam search2025.12 | 90.48 | 4.74 | 91.23 | 13.89 | 95.4 | 11.31 | |
| LMSanitatorFuzzy Search Rounds (FS)=1002025.12 | 90.48 | 20.12 | 91.81 | 103.6 | 92.53 | 83.65 | |
| Patronus-greedyFuzzy Search Rounds (FS)=3, Optimization Strategy=greedy optimization2025.12 | 88.1 | 1.12 | 81.29 | 2.42 | 90.23 | 2.44 | |
| LMSanitatorFuzzy Search Rounds (FS)=502025.12 | 88.1 | 9.1 | 81.87 | 34.43 | 92.53 | 29.98 | |
| LMSanitatorFuzzy Search Rounds (FS)=252025.12 | 84.52 | 4.11 | 70.18 | 12.29 | 83.33 | 12.2 | |
| Patronus-greedyFuzzy Search Rounds (FS)=2, Optimization Strategy=greedy optimization2025.12 | 83.33 | 0.74 | 71.35 | 1.6 | 85.63 | 1.54 |