Generating adversarial examples for DNN using pooling layers

Issue title: Special Section: Fuzzy theoretical model analysis for signal processing

Guest editors: Valentina E. Balas, Jer Lang Hong, Jason Gu and Tsung-Chih Lin

Article type: Research Article

Authors: Zhang, Yueling^{a; *} | Pu, Geguang^a | Zhang, Min^a | Y, William^b

Affiliations: [a] Computer Science and Software Engineering Institute, East China Normal University, Shanghai, China | [b] Department of Computer Science, Princeton University, Princeton, NJ, USA

Correspondence: [*] Corresponding author. Yueling Zhang, Computer Science and Software Engineering Institute, East China Normal University, 3663 North Zhongshan Rd, 200062, Shanghai, China. E-mail: ylzhang.ecnu@gmail.com.

Abstract: Deep Neural Network is an application of Big Data, and the robustness of Big Data is one of the most important issues. This paper proposes a new approach named PCD for computing adversarial examples for Deep Neural Network (DNN) and increase the robustness of Big Data. In safety-critical applications, adversarial examples are big threats to the reliability of DNNs. PCD generates adversarial examples by generating different coverage of pooling functions using gradient ascent. Among the 2707 input images, PCD generates 672 adversarial examples with L∞ distances less than 0.3. Comparing to PGD (state-of-art tool for generating adversarial examples with distances less than 0.3), PCD finds 1.5 times more adversarial examples than PGD (449) does.

Keywords: Deep neural network, robustness, coverage, big data

DOI: 10.3233/JIFS-179295

Journal: Journal of Intelligent & Fuzzy Systems, vol. 37, no. 4, pp. 4615-4620, 2019