Hue and Saturation oriented Imperceptible Adversarial Examples and their Evaluation

With the advancement of machine learning technology, AI can be applied to various systems such as autonomous driving and medical diagnosis. On the other hand, the vulnerability of AI system is pointed out. The adversarial examples (AEs), which causes incorrect decisions by AI, is one of the threats for AI. For using AI safely, AEs has to be investigated thoroughly. In order to evaluate threat of AEs, this paper proposes the generating method for AEs which is based on hue and saturation property of the image. The proposed method which is based on the conventional gradient-based method, introduces constraints for projection in the HSV color space. The proposed method utilizes properties of human vision which is insensitive to changes in color and sensitive to changes in brightness. These properties contribute imperceptibility of AEs. Experiments revealed the vulnerability of DNNs against imperceptible AEs and clarified the importance of countermeasures.