Adversarial Defense with Local Robust Principal Component Analysis and Wavelet Denoising

In this work, a local robust principal component analysis (RPCA)-based adversarial defense methodology is presented. The proposed defense is inspired by recent works of input reconstruction paradigm. The proposed method uses localised processing to apply RPCA towards removal of adversarial perturbations. The reconstructed image is guided by applying wavelet denoising to remove residual perturbation which improves the defense more efficiently. Experimental results are presented using state-of-the-art C&W attack and Square attack to show the effectiveness of the proposed adversarial defense on a pre-trained classifier model.

I. Introduction

Over the recent years, Deep Neural Networks (DNNs) have become progressively popular and successful in various signal and image processing applications. Although these state-of-the-art DNNs are utilized for detection, segmentation, and classification etc., the presence of specially crafted adversarial perturbations pose as a major challenge towards the robust and accurate functioning of the deep networks. Such images, called as adversarial examples, have piqued the interest of the researchers over the recent years towards reducing the vulnerability of the DNNs and thus, improving the overall system performance.