Table 10 Performance of NHANet model with different epsilons under BIM attack algorithm.

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

 

Original

\(\upepsilon\) = 1/255

\(\upepsilon\) = 2/255

\(\upepsilon\) = 3/255

\(\upepsilon\) = 4/255

\(\upepsilon\) = 5/255

Adversarial accuracy

0.9840

0.8760

0.4773

0.2133

0.1227

0.0840

Attack success rate

0.0

0.1098

0.5149

0.7832

0.8753

0.9146

Average L1 distance

0.0

1.1665

2.2306

3.2206

4.0604

4.9083

Average L2 distance

0.0

0.0676

0.1307

0.1914

0.2445

0.2995

Back to article page