Adversarial UFP/UFN Attack Evolution

In this paper, we explore the effects of using evolutionary computation in an effort to increase the effectiveness of universal false positive and universal false negative attacks. These attacks are focused on machine learning based fake news detection systems. For this task, we use a steady-state genetic algorithm to evolve a set of adversarial samples. Our results suggest that using a steady-state genetic algorithm to evolve adversarial samples yields a noticeable increase in misclassification rates and exposes potential vulnerabilities in the evaluated machine learning algorithms.