Visual Quality Assessment of Adversarially Attacked Images

Ali H. Mezher, Yingpeng Deng, Lina J. Karam

Abstract: With the growing interest in developing more effective and less perceivable adversarial attacks on images, many existing objective visual quality assessment metrics are being used in order to benchmark the performance of these attacks in terms of their imperceptibility. However, these objective metrics were originally developed for quantifying the image quality degradations due to other factors, such as acquisition, compression, and transmission, and there is no clear consensus on the effectiveness of these metrics in evaluating adversarial attacks, nor what the best metrics for such attacks are. This work proposes a subjective study where human subjects rate the visual quality of images that have been adversarially attacked using popular and contemporary adversarial attack methods, compared to their original form. The collected subjective scores along with the original and adversarially attacked images are used to construct a Subjective Testing Adversarial Attack Quality (STAAQ) database. The constructed database is used to assess the performance of popular objective metrics in terms of their ability to predict the visual quality of adversarially attacked images.