Deep Learning for Malware Classification

Abstract

Malware detection tasks have traditionally been solved using hand-crafted features obtained through heuristic processes by experts. Current research suggests that deep convolutional neural networks can excel for automated feature extraction from raw data inputs. However, malware activities occur from the transference of executable code blocks which perform individual functionalities, and capturing such temporal movements is fundamental for successfully detecting malware. The success of recurrent neural networks for sequential data, such as speech and language applications inspires us to propose a generic deep framework for malware detection based on convolutional and recurrent neural networks which: (i) does not require domain knowledge for feature design and (ii) explicitly exploits spatial and temporal characteristics in input data. We evaluate our framework and show that it outperforms the traditional approach.