[3Rin4] Interactive 1

[3Rin4-69] Visualization and observation of dragging class transition induced by targeted adversarial attacks

〇Kense Todo¹, Shoya Yasuda¹, Masayuki Yamamura¹ (1.Tokyo Institute of Technology School of Computing)

Keywords:adversarial examples, Genetic algorithm, Understanding behavior

Adversarial examples in machine learning have got more attention these days. Making “universal” “targeted” attacks against “black-boxed” neural networks is slightly difficult, therefore structural analysis of neural networks using adversarial attacks is not yet sufficient. In this work, we created image noises for universal, targeted and black-boxed attacks by genetic algorithm, and applied to another class samples to investigate their transition behavior in the compressed feature space. As a result, a targeted noise for one class induced a characteristic transition (“dragging”) for another class samples. Since the degree of dragging differed in classes, each class seems to have different resistance against adversarial attacks. This could be some indicator corresponding to the distance between classes and would be a clue to the analysis of the feature structure of black-boxed neural networks.