Websites utilize several approaches to detect automated agents. The agents are deployed either for beneficial purposes such as search engine crawlers, or to perform tasks on behalf of the adversary such as scanning for vulnerabilities. Recent methods in detecting such agents include the analysis of the behavior that the agents show when visiting the website. In this paper, I) we describe a deep learning framework that analyzes the triggered browser events to classify the visitor. II) We develop two adversarial attacks in order to bypass the defense by generating adversarial vectors that are misclassified by the model. III) We discuss how applicable the attacks are by reviewing the limitations of the popular tools (i.e., Selenium and Puppeteer) used for the development of automated agents based on full-fledged browsers.
