A Semi-supervised Learning Methodology for Malware Categorization using Weighted Word Embeddings

Due to the vertiginous growth of malicious actors, malware has been crafted, distributed and propagated around the world with new and sophisticated techniques. Classical malware detection procedures, mostly based on signatures and heuristic searches, are now being replaced with machine learning-based (ML) solutions. However, some challenges are still present. Firstly, supervised approaches use anti-virus tags to create hand-crafted datasets, resulting in a lack of taxonomy and uncertainty if a given observation is classified with a proper label. Secondly, off-line and feed-forward approaches may result in complex and time consuming feature extraction tasks. In this work, we propose a novel method that reinforces malware characterization by capturing rich relevance and contextual patterns into an n-dimensional weighted word embedding vector (WEV) space. Results prove that by clustering similar WEVs via unsupervised learning, malware can be categorized into four major families, improving detection with less resources.