Deep Combination of Stylometry Features in Forensic Authorship Analysis

Abstract

Authorship Analysis (AA) in forensic is a process aim to extract information about an author from his/her writings.Forensic AA is needed for detection characteristics of anonymous authors to make better the security of digital media userswho are exposed to disturbing entries such as threats or harassment emails. To analyze whether two anonymous short textswere written by the same author, we propose a combination of stylometry features from different categories in differentprogress. In the majority of the previous AA studies, the stylometric features from different categories are concatenated in apreprocess. In these studies, during the learning process, no category-specific operations are performed; all categories used areevaluated equally. On the other hand, the proposed approach has a separate learning process for each feature category due totheir qualitative and quantitative characteristics and combines these processes at the decision phase by using a Combination ofDeep Neural Networks (C-DNN). To evaluate the Authorship Verification (AV) performance of the proposed approach, wedesigned and implemented a problem-specific Deep Neural Network (DNN) for each stylometry category we used.Experiments were conducted on two English public datasets. The results show that the proposed approach significantlyimproves the generalization ability and robustness of the solutions, and also have better accuracy than the single DNNs.