Homoglyphs restoration with deep learning - focus on optical character recognition -

Abstract

Homoglyphs are a shape that is difficult to distinguish visually because it is similar or identical. Because of this characteristic, attackers use them for phishing, causing serious problems. In this paper, we deal with the countermeasures against a new type of attacks through homoglyphs. Existing homoglyph attacks consist of characters or words, and so do the corresponding countermeasures. However, since the new type contains multiple homoglyphs in the text of the scam email, a new countermeasure was needed. To this end, we present a natural language processing technique that utilizes sequence information and a method that utilizes both visual elements and sequence information compared to existing methods for restoring homoglyphs. We use accuracy and false-positive rate (FPR) as evaluation criteria to compare the existing methods with the newly proposed methods. Through comparison using multiple evaluation criteria, we show that the method using both visual judgment and sequence information converts the homoglyph most accurately.