Deobfuscating leetspeak with deep learning to improve spam filtering

Abstract

The evolution of anti-spam filters has forced spammers to make greater efforts to bypass filters in order to distribute content over networks. The distribution of content encoded in images or the use of Leetspeak are concrete and clear examples of techniques currently used to bypass filters. Despite the importance of dealing with these problems, the number of studies to solve them is quite small, and the reported performance is very limited. This study reviews the work done so far (very rudimentary) for Leetspeak deobfuscation and proposes a new technique based on using neural networks for decoding purposes. In addition, we distribute an image database specifically created for training Leetspeak decoding models. We have also created and made available four different corpora to analyse the performance of Leetspeak decoding schemes. Using these corpora, we have experimentally evaluated our neural network approach for decoding Leetspeak. The results obtained have shown the usefulness of the proposed model for addressing the deobfuscation of Leetspeak character sequences.