no code implementations • 22 Jul 2021 • Gihyuk Ko, Gyumin Lim
From this insight, we propose an unsupervised detection of adversarial examples using reconstructor networks trained only on model explanations of benign examples.
