Learning to Discriminate Adversarial Examples by Sensitivity Inconsistency in IoHT Systems

Table 7

Detection performance of different features. None means features dimensionality is 300 for IMDB and 100 for AG’s news; in top settings, for IMDB and for AG’s news.

Top IWs	Sensitivity flags	Softmax distribution	IMDB		AG’s news
			Recall (%)	F1-score (%)	Recall (%)	F1-score (%)
None	✓		80.0	89.5	78.6	72.3
		✓	92.4	94.1	84.9	87.6
	✓	✓	92.0	91.5	90.1	90.5
	✓		84.4	92.4	76.9	77.5
		✓	95.2	96.1	89.4	90.7
	✓	✓	99.8	97.7	95.5	91.5

Bold values indicate the best results in different feature settings.