HVS-inspired adversarial image generation with high perceptual quality

Abstract

Adversarial images are able to fool the Deep Neural Network (DNN) based visual identity recognition systems, with the potential to be widely used in online social media for privacy-preserving purposes, especially in edge-cloud computing. However, most of the current techniques used for adversarial attacks focus on enhancing their ability to attack without making a deliberate, methodical, and well-researched effort to retain the perceptual quality of the resulting adversarial examples. This makes obvious distortion observed in the adversarial examples and affects users’ photo-sharing experience. In this work, we propose a method for generating images inspired by the Human Visual System (HVS) in order to maintain a high level of perceptual quality. Firstly, a novel perceptual loss function is proposed based on Just Noticeable Difference (JND), which considered the loss beyond the JND thresholds. Then, a perturbation adjustment strategy is developed to assign more perturbation to the insensitive color channel according to the sensitivity of the HVS for different colors. Experimental results indicate that our algorithm surpasses the SOTA techniques in both subjective viewing and objective assessment on the VGGFace2 dataset.