An Approximately Optimal Bot for Non-Submodular Social Reconnaissance

Abstract

The explosive growth of Online Social Networks in recent years has led to many individuals relying on them to keep up with friends & family. This, in turn, makes them prime targets for malicious actors seeking to collect sensitive, personal data. Prior work has studied the ability of socialbots, i.e. bots which pretend to be humans on OSNs, to collect personal data by befriending real users. However, this prior work has been hampered by the assumption that the likelihood of users accepting friend requests from a bot is non-increasing -- a useful constraint for theoretical purposes but one contradicted by observational data. We address this limitation with a novel curvature based technique, showing that an adaptive greedy bot is approximately optimal within a factor of 1 - 1/e1/δ ~0.165. This theoretical contribution is supported by simulating the infiltration of the bot on OSN topologies. Counter-intuitively, we observe that when the bot is incentivized to befriend friends-of-friends of target users it out-performs a bot that focuses on befriending targets.