[4J3-GS-2] Machine learning: Adversarial examples and security

[4J3-GS-2-03] An Attempt of Evolutionary Generation of Adversarial Examples for One-Shot Depth Estimation

〇Renya Daimo¹, Takahiro Suzuki¹, Satoshi Ono¹ (1. Kagoshima University)

Keywords:One-Shot Depth Estimation, Adversarial example, Evolutionary computation, Deep Neural Network

With recent advances of Deep Neural Networks(DNNs), the performance of monocular depth estimationhas been improved and expectations for its practical use are increasing. On the other hand, recent studies haverevealed vulnerabilities of DNNs, in which carefully designed perturbations called adversarial examples can causemisclassi cation. It is essential to investigate such vulnerabilities so that DNN-based one-shot depth estimatorscan be safely applied to real-world applications. Therefore, this study proposes an evolutionary computation-based method to generate adversarial examples that cause one-shot depth estimators to measures make incorrectmeasurements. The proposed method performs targeted attack under black-box condition by adding perturbationson a target object so that a target object would not be detected. Experimental results demonstrated that theproposed method successfully generated AEs that misleaded a well-known DNN-based one-shot depth estimator.