Learning to Discriminate Adversarial Examples by Sensitivity Inconsistency in IoHT Systems

Table 6

Generalization evaluation in the toughest scenario. The values in parentheses are default effect. Bold values denotes that the detection performance under transferability is better than the default effect.

Dataset	Model	PWWS		BAE		Deepwordbug
		Recall (%)	F1-score (%)	Recall (%)	F1-score (%)	Recall (%)	F1-score (%)
Yelp	LSTM	86.2 (91.6)	85.7 (90.5)	90.6 (94.5)	91.5 (93.2)	92.8 (91.2)	89.7 (87.4)
	RoBERTa	87.4 (94.7)	89.6 (92.6)	87.3 (92.0)	87.9 (92.7)	85.9 (92.8)	87.0 (89.1)
AG’s news	LSTM	90.3 (94.4)	87.1 (88.5)	93 (91.7)	93.5 (90.3)	91.4 (88.6)	89.2 (84.1)
	RoBERTa	90.8 (91.3)	87.6 (85.4)	87.0 (89.9)	85.7 (91.2)	88.9 (86.6)	90.0 (87.5)