Adversarial Machine Learning in the Wild

Abstract

Deep neural networks are making their way into our everyday lives at an increasing rate. While the adoption of these models has greatly improved our everyday lives, it has also opened the door to new vulnerabilities in real-world systems. More specifically, in the scope of this work we are interested in one class of vulnerabilities: adversarial attacks. Given the high importance and the sensitivity of some of the tasks these models are responsible for, it is crucial to study such vulnerabilities in real-world systems. In this work, we look at examples of deep neural network-based real-world systems, vulnerabilities of such systems, and approaches for making such systems more robust. First, we study an example of leveraging a deep neural network in a business-critical real-world system. We discuss how deep neural networks improve the quality of smart voice assistants. More specifically, we introduce how collaborative filtering models can automatically detect and resolve the errors of a voice assistant. We then discuss the success of this approach in improving the quality of a real-world voice assistant. Second, we demonstrate a proof of concept for an adversarial attack against content-based recommendation systems which are commonly used in real-world settings. We discuss how malicious actors can add unnoticeable perturbations to the content they upload to the website to achieve their preferred outcomes. We also show how adversarial training can render such attacks useless. Third, we discuss another example of how adversarial attacks can be leveraged to manipulate a real-world system. We study how adversarial attacks can successfully manipulate YouTube's copyright detection model and the financial implications of this vulnerability. In particular, we show how adversarial examples created for a copyright detection model that we implemented transfer to another black-box model. Finally, we study the problem of transfer learning in an adversarially robust setting. We discuss how robust models contain robust feature extractors and how we can leverage them to train new classifiers that preserve the robustness of the original model. We then study the case of fine-tuning in the target domain while preserving the robustness. We show the success of our proposed solutions in preserving the robustness in the target domain.