Scientific Reports

Table 11 Performance of various models under different epsilons.

From: An incremental adversarial training method enables timeliness and rapid new knowledge acquisition

Neural network model	Evaluation metrics	\({\varvec{\upepsilon}}\)(epsilons)
Neural network model	Evaluation metrics	original	1/255	2/255	3/255	4/255	5/255
DNN	Adversarial accuracy	0.9627	0.8893	0.4907	0.1960	0.1133	0.0933
	Attack success rate	0.0	0.0762	0.4903	0.7964	0.8823	0.9030
	Average L₁ distance	0.0	1.1702	2.2979	3.3012	3.8961	4.2859
	Average L₂ distance	0.0	0.0677	0.1338	0.1937	0.2355	0.2715
BiLSTM	Adversarial accuracy	0.9640	0.8440	0.3987	0.1667	0.0920	0.0693
	Attack success rate	0.0	0.1245	0.5864	0.8271	0.9046	0.9281
	Average L₁ distance	0.0	1.1719	2.3116	3.3308	3.9296	4.3159
	Average L₂ distance	0.0	0.0678	0.1343	0.1949	0.2369	0.2730
CNN-LSTM	Adversarial accuracy	0.9560	0.6160	0.1933	0.1307	0.1067	0.1027
	Attack success rate	0.0	0.3556	0.7978	0.8633	0.8884	0.8926
	Average L₁ distance	0.0	1.1421	2.1842	3.0447	3.6027	4.0221
	Average L₂ distance	0.0	0.0666	0.1291	0.1832	0.2231	0.2587
BiLSTM-MultiheadAttention	Adversarial accuracy	0.9587	0.6120	0.1760	0.0853	0.0667	0.0613
	Attack success rate	0.0	0.3616	0.8164	0.9110	0.9305	0.9360
	Average L₁ distance	0.0	1.1703	2.2859	3.2548	3.8348	4.2278
	Average L₂ distance	0.0	0.0677	0.1334	0.1919	0.2330	0.2687
CNN-LSTM-ATT	Adversarial accuracy	0.9307	0.4107	0.1573	0.1040	0.0880	0.0747
	Attack success rate	0.0	0.5587	0.8309	0.8883	0.9054	0.9198
	Average L₁ distance	0.0	1.1120	2.0774	2.8439	3.3736	3.8137
	Average L₂ distance	0.0	0.0657	0.1248	0.1746	0.2130	0.2483

Back to article page

Search

Advanced search

Quick links