Detect and defense against adversarial examples in deep learning using natural scene statistics and adaptive denoising

Despite theenormous performance of deep neural networks (DNNs), recent studies have showntheir vulnerability to adversarial examples (AEs), i.e., carefully perturbedinputs designed to fool the targeted DNN. Currently, the literature is richwith many effective attacks to craft such AEs. Meanwhile, many defensestrategies have been developed to mitigate this vulnerability. However, theselatter showed their effectiveness against specific attacks and does notgeneralize well to different attacks. In this paper, we propose a framework fordefending DNN classifier against adversarial samples. The proposed method isbased on a two-stage framework involving a separate detector and a denoisingblock. The detector aims to detect AEs by characterizing them through the useof natural scene statistic (NSS), where we demonstrate that these statisticalfeatures are altered by the presence of adversarial â€¦