Research Article
Learning to Discriminate Adversarial Examples by Sensitivity Inconsistency in IoHT Systems
Table 5
Transferability evaluation of different models. Bold values indicate the better scores among two target models that model
migrate to.
| Model | TextFooler | PWWS | BAE | Deepwordbug | Recall (%) | F1-score (%) | Recall (%) | F1-score (%) | Recall (%) | F1-score (%) | Recall (%) | F1-score (%) |
| CNN | 99.2 | 96.3 | 98.1 | 95.2 | 98.2 | 95.8 | 98.6 | 95.5 | LSTM | 82.6 | 86.8 | 86.6 | 84.1 | 88.4 | 90.9 | 81.8 | 89.6 | BERT | 89.2 | 93.8 | 91.6 | 92.6 | 90.4 | 88.7 | 89.7 | 92.5 |
| CNN | 98.5 | 95.1 | 99.9 | 94.6 | 97.5 | 94.8 | 99.4 | 93.7 | LSTM | 97.5 | 95.4 | 92.1 | 92.2 | 96.5 | 95.5 | 92.3 | 91.6 | BERT | 97.2 | 95.7 | 95.0 | 95.2 | 91.8 | 93.3 | 94.8 | 94.7 |
| CNN | 99.1 | 96.1 | 100.0 | 93.7 | 97.1 | 96.3 | 98.7 | 96.0 | LSTM | 95.4 | 95.2 | 92.2 | 91.4 | 96.2 | 95.5 | 95.4 | 94.7 | BERT | 97.0 | 97.0 | 97.4 | 96.5 | 95.5 | 95.5 | 95.9 | 95.9 |
|
|