Preventing Hidden Information Leaks Using Author Attribution Methods and Neural Networks
Alexander Khazagarov, Alisa A. Vorobeva, Viktoriia Mikhailovna Korzhuk · 2021
This paper addresses the problem of hidden information leakage detection through the use of text steganography. Presented comparative research results show how to perform this task by detecting changes in user's writing styles using neural networks and various types of text features. The framework for hidden leakages detection based on discovering changes in the author's writing style with deep neural networks (RNN, LSTM, GRU, CNN) is proposed. A series of experiments on text corpus containing Russian online texts were carried out to evaluate hidden leakages detection accuracy. The experiments showed that the LSTM and character 4-grams together allow achieving the accuracy of 87%. Text preprocessing significantly decreases accuracy which is also shown.