End to End Autorship Email Verification Framework for a Secure Communication

The paper proposes an alternative email account protection to prevent a very specific targeting email attacks where an attacker can impersonate a legitimate/trusted sender to steal personal information to the recipient. Authorship mechanism based on the analysis of the author's writing style and implemented through binary traditional and deep learning classifiers is applied to build the email verification mechanism. A flexible architecture, where the authorship component can be placed in different locations, is proposed. Due to its location and consequently to the email data available, can be exploited an individual writing style, or an end to end writing style learning related to the sender-receiver communication. The system is validated on two different dataset (i) the well-known public Enron dataset, with the experiments showing the author verification accuracy of 96.5% and 99% respectively for the individual and end to end writing style learning and (ii) our private dataset, with accuracy results of 98.3% and 97%. An alternative classification training, that exploits the partition of the dataset in subsets having approximately the same length, is presented. From the results obtained is proved how such training approach outperforms the traditional training where emails of different lengths are contained in the same training dataset. The overall results obtained proved that the authorship mechanism proposed is a promising alternative support technique exploitable as an email anti-scam or anti-theft tool to guarantee secure email communication.