Multi-Class Triplet Loss With Gaussian Noise for Adversarial Robustness

Abstract

Deep Neural Networks (DNNs) classifiers performance degrades under adversarial attacks, such attacks are indistinguishably perturbed relative to the original data. Providing robustness to adversarial attacks is an important challenge in DNN training, which has led to extensive research. In this paper, we harden DNN classifiers under the adversarial attacks by regularizing their deep internal representation space with Multi-class Triplet regularization method. This method enables DNN classifier to learn a feature representation that detects similarities between adversarial and clean images and brings similar images close to their original class and pushes dissimilar images away from their false classes. This training process with our Multi-class Triplet regularization method in combination with Gaussian noise injection proves to be more robust in detecting adversarial attacks exceeding that of adversarial training on strong iterative attacks.

Dynamic Programming-based Adversarial Windows Payload Generator

“…Adversarial attacks have been investigated in the area of image, audios, texts and recently in Windows executable files classification exercise and a number of successful attacks examples in image, audios and texts have been generated to cause misclassification [1][2][3][4][5][6][7]. The principal reason for the success in image, audios and texts is that their feature-space is comparatively fixed, an image or text can be formatted as a three-dimensional array of pixels with each pixel value as a three-dimensional RGB (red, green, blue) vector value ranged from 0 to 255, thus, is feasible to find an exact function that is differentiable, therefore, a feature-space attack built on gradients can instantly apply on text or images to create adversarial attack examples.…”