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