Open AccessA Two-layer Mechanism Identification Method for VIN Adversarial Examples
Author(s) -
Yingdi Wang,
Jiqiang Liu,
Tong Chen,
Yingxiao Xiang,
Endong Tong,
Wenjia Niu
Publication year - 2020
Publication title -
journal of physics. conference series
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 0.21
H-Index - 85
eISSN - 1742-6596
pISSN - 1742-6588
DOI - 10.1088/1742-6596/1584/1/012069
Subject(s) - adversarial system , computer science , identification (biology) , reinforcement learning , artificial intelligence , path (computing) , feature (linguistics) , domain (mathematical analysis) , layer (electronics) , machine learning , key (lock) , mechanism (biology) , mathematics , computer security , mathematical analysis , linguistics , philosophy , botany , chemistry , organic chemistry , epistemology , biology , programming language
With the rapid development of machine learning algorithms, the security problems are gradually emerging. Most existing machine learning algorithms may be attacked by adversarial examples. Especially in the domain of path planning, the adversarial maps may result in multiple harmful consequences on the paths predicted by Deep Reinforcement Learning (DRL) algorithms. However, there is no suitable approach to automatically identify them. To our knowledge, all previous work used manual observation method to identify the adversarial maps, which is time-consuming. Therefore, this paper explores a two-layer mechanism method to automatically identify the adversarial examples in Value Iteration Networks (VIN). We define the four categories of attack results and identify them by combining path feature comparison and path image classification. Experiments show that this method can achieve an effective identification on VIN adversarial examples.