Table 10 Comparison on attacking WordLSTM trained with adversarial training. The optimal value of each metric is emphasized in bold.
From: Hard label adversarial attack with high query efficiency against NLP models
Data | Metric | Attack | ||||
|---|---|---|---|---|---|---|
HLBB | TextHoaxer | LeapAttack | SSPAttack | QEAttack | ||
AG | Sim.% | 90.13 | 88.45 | 87.97 | 90.20 | 91.06 |
Pert.% | 12.37 | 12.59 | 13.45 | 9.84 | 9.39 | |
Suc.% | 90.24 | 90.34 | 90.24 | 90.34 | 90.34 | |
IMDB | Sim.% | 94.68 | 94.96 | 94.65 | 93.98 | 95.47 |
Pert.% | 4.96 | 4.21 | 4.33 | 3.97 | 3.46 | |
Suc.% | 90.73 | 90.73 | 90.73 | 90.73 | 90.73 | |
