Reverse Engineering the Behaviour of Twitter Bots

摘要

Recent research has shown significant success in the detection of social bots. While there are tools to distinguish automated bots from regular user accounts, information about their strategies, biases and influence on their target audience remains harder to obtain. To uncover such details, e.g., to understand the role of bots in political campaigns, we address three questions: Can we describe the behaviour of a bot (when and how a bot takes actions) by a set of understandable rules?How can we express bias and influence? Can we extract such information automatically, from observations of a bot?In this paper, we present an approach to reverse engineering the behaviour of Twitter bots to create a visual model explaining their actions. We use machine learning to infer a set of simple and general rules governing the behaviour of a bot. We propose the notion of differential sentiment analysis to provide means of understanding the behaviour with respect to the topics on its network in relation to both its sources of information (friends) and its target audience (followers). Respectively, this provides insights into their bias and the influence aimed at their target audience. We evaluate our approach using prototype bots we created and selected real Twitter bots. The results show that we are successful in correctly describing the behaviour of the bots and potentially useful in understanding their impact.