Siamese neural network architecture for homoglyph attacks detection

Abstract

Abstract Primarily an adversary uses homoglyph or spoofing attack approach to obfuscate domain name, file name or process names. This approach facilitates to create domain name, file name or process names which look visually homogeneous to legitimate domain name, file name or process names. This paper introduces Siamese neural network architecture which uses the application of recurrent structures with Keras character level embedding to learn the optimal features by considering an input in the form of raw strings. For comparative study, various recurrent structures are used. The performances obtained by recurrent structures are almost closer. However, the proposed method performed well in comparison to the existing methods such as Edit Distance, Visual Edit Distance and Siamese convolutional neural networks.