A Method Generating Adversarial Mark Based on Convolutional Neural Networks

Deep neural networks (DNNs) haveZhengjie, D.Meijun, L. achieved impressive results in image classification tasks. Recent studies have shown that adding imperceptible perturbations to original images can cause image recognition models to make erroneous judgments. Additionally, visible marks, which add text or image information to images, serve as a reminder of copyright ownership and can help preventing image theft and infringement. In a sense, visible text mark can be seen as meaningful noise added to clean images. In this article, we propose a method (GAM) that combines text mark and adversarial examples by focusing perturbations on meaningful text that does not affect human judgment, while causing image classification models to produce incorrect results.