A novel generalized adversarial image method using descriptive features

Abstract

Abstract— Currently, machine learning not only solves simple problems such as object classification but also machine learning is widely applied in the field of computer vision such as identification systems, object detection, and modules in the authentication system, intelligent processing algorithms such as automatic driving, chatbot, etc. Deep learning models on GAN networks can automatically generate image data such as objects, animals, human faces, or the like by learning the word features of images in datasets such as MS-COCO, ImageNET, CUB, etc. Using this technique, attackers can fake images in some cases with malicious intent. In this paper, the authors propose to build a Generative Adversarial Network to create images that fool the target model YOLOv7, INCEPTIONv3. Experimental results on the CUB dataset show our proposed model's ability to generate adversarial examples is highly effective with an average image generation time equal to 0.16 seconds/an image. The successful rate of fooling the model reached over 85%, average recognition rate reached over 45% for the YOLOv7 model. When experimenting on the INCEPTIONv3 model, the successful rate of fooling the model reached over 95%, average recognition rate reached over 50%. The image fidelity evaluated by the PSNR index reached an average of greater than 29.