Efficiently Detecting Web Spambots in a Temporally Annotated Sequence

Abstract

Web spambots are becoming more advanced, utilizing techniques that can defeat existing spam detection algorithms. These techniques include performing a series of malicious actions with variable time delays, repeating the same series of malicious actions multiple times, and interleaving legitimate (decoy) and malicious actions. Existing methods that are based on string pattern matching are not able to detect spambots that use these techniques. In response, we define a new problem to detect spambots utilizing the aforementioned techniques and propose an efficient algorithm to solve it. Given a dictionary of temporally annotated sequences $\hat{S}$ modeling spambot actions, each associated with a time window, a long, temporally annotated sequence T modeling a user action log, and parameters f and k, our problem seeks to detect each sequence in $\hat{S}$ that occurs in T at least f times within its associated time window, and with at most k mismatches. Our algorithm solves the problem exactly, it requires linear time and space, and it employs advanced data structures and the Kangaroo method, to deal with the problem efficiently.