Comparison of LSTM Architecture for Malware Classification

An experimental study of malware and benign classification from Windows API call sequences dataset using a deep learning framework is presented. We conduct a series of Long Short-Term Memory (LSTM) modifications, Bidirectional Long Short-Term Memory (BiLSTM). The proposed one architecture, such a half per half input sequence processed on the Siamese BiLSTM network looks. All three base models are treated fairly with scenario series of modification such a callback, batch normalization, dropout, and attention mechanism. As the results of this experiment, adding dropout and attention mechanisms show improvement from baseline models. In addition, we find that our proposed architecture with dropout and attention mechanism slightly outperform from other models.