Authorship Attribution With Few Training Samples

Abstract

This chapter discusses authorship attribution through a training sample. The focus on authorship attribution discussed in this chapter differs in two ways from the traditional authorship identification problem discussed in the earlier chapters of this book. Firstly, the traditional authorship attribution studies [63, 65] only work in the presence of large training samples from each candidate author, which are typically enough to build a classification model. With authorship attribution, the emphasis is on using a few training samples for each suspect. In some scenarios, no training samples may exist, and the suspects may be asked (usually through court orders) to produce a writing sample for investigation purposes. Secondly, in traditional authorship studies, the goal is to attribute a single anonymous document to its true author. In this chapter, we look at cases where we have more than one anonymous message that needs to be attributed to the true author(s). It is likely that the perpetrator may either create a ghost e-mail account or hack an existing account, and then use it for sending illegitimate messages in order to remain anonymous. To address the aforementioned shortfalls, the authorship attribution problem has been redefined as follows: given a collection of anonymous messages potentially written by a set of suspects {S₁, ···, S_n}, a cybercrime investigator first wants to identify the major groups of messages based on stylometric features; intuitively, each message group is written by one suspect. Then s/he wants to identify the author of each anonymous message collection from the given candidate suspects. To address the newly defined authorship attribution problem, the stylometric pattern-based approach of AuthorMinerl (described previously in Sect. 5.4.1) is extended and called AuthorMinerSmall. When applying this approach, the stylometric features are first extracted from the given anonymous message collection Ω.