Detect and Defense Against Adversarial Examples in Deep Learning using Natural Scene Statistics and Adaptive Denoising

Despite the enormous performance of deepneural networks (DNNs), recent\nstudies have shown theirvulnerability to adversarial examples (AEs), i.e.,\ncare-fully perturbed inputs designed to fool the targetedDNN. Currently, the\nliterature is rich with many ef-fective attacks to craft such AEs. Meanwhile,\nmany de-fenses strategies have been developed to mitigate thisvulnerability.\nHowever, these latter showed their effec-tiveness against specific attacks and\ndoes not general-ize well to different attacks. In this paper, we proposea\nframework for defending DNN classifier against ad-versarial samples. The\nproposed method is based on atwo-stage framework involving a separate detector\nanda denoising block. The detector aims to detect AEs bycharacterizing them\nthrough the use of natural scenestatistic (NSS), where we demonstrate that\nthese statis-tical features are altered by the presence of\nadversarialperturbations. The denoiser is based on block matching3D (BM3D)\nfilter fed by an optimum threshold valueestimated by a convolutional neural\nnetwork (CNN) toproject back the samples detected as AEs into theirdata\nmanifold. We conducted a complete evaluation onthree standard datasets namely\nMNIST, CIFAR-10 andTiny-ImageNet. The experimental results show that\ntheproposed defense method outperforms the state-of-the-art defense techniques\nby improving the robustnessagainst a set of attacks under black-box, gray-box\nand white-box settings. The source code is available at:\nhttps://github.com/kherchouche-anouar/2DAE\n