Minimum Noticeable Difference-Based Adversarial Privacy Preserving Image Generation

Deep learning models are found to be vulnerable adversarial examples, as wrong predictions can caused by small perturbation in input for deep models. Most of the existing works image generation try achieve attacks most models, while few them make efforts on guaranteeing perceptual quality examples. High examples matter many applications, especially privacy preserving. In this work, we develop a framework based Minimum Noticeable Difference (MND) concept generate preserving images that have minimum difference from clean ones but able attack To this, an loss is firstly proposed attacked successfully. Then, quality-preserving developed taking magnitude and perturbation-caused structural gradient changes into account, which aims preserve high generation. best our knowledge, first work exploring MND evaluate its performance terms quality, classification face recognition tested with method several anchor methods work. Extensive experimental results demonstrate capable generating remarkably improved metrics (e.g., PSNR, SSIM, MOS) than generated methods.