Improving robust accuracy through gradient boosting with ADP

Abstract

In adversarial examples, humans can easily classify the images even though the images are corrupted. However, recently, some related work has shown that deep neural networks are vulnerable to adversarial attacks [1]. To increase the robustness against adversarial attacks, many methods were carried out, such as k-Winners [2], robust sparse Fourier Transform [3], and Compact Convolution [4]. Many of the defense strategies aimed to mark the gradient, train different classifiers, and use new loss calculations. In the thesis, several ensemble models were trained by applying both typical gradient boosting and enlarging the diversity among base models to improve their robustness against adversarial attacks. The purpose is to show that making adversarial examples difficult to transfer among individual members would cause the state-of-the-art attacking algorithms to fail to attack the trained robust ensemble model to a certain extent.