Learning from the Environment: A Novel Adversarial Patch Attack against Object Detectors Using a GAN Trained on Image Slices

Over the last few decades, object detection technology leveraging Deep Neural Networks (DNNs) has made significant progress and has found extensive applications in diverse areas, including surveillance systems. However, its potential for malicious use poses significant privacy risks. In this work, a novel method is proposed in order to generate adversarial patches to degrade the performance of malicious DNNs. Our approach involves training a Generative Adversarial Network (GAN) on dataset slices to learn background visual features. The optimal adversarial patch is then searched among the images generated by the GAN. The adversarial patches produced through this method exhibit improved attack efficacy and environmental consistency. By conducting comprehensive experiments in digital as well as physical settings, we show the performance of our method in comparison to baseline approaches.