Data Augmentation for Code Analysis

Abstract

A key challenge of applying machine learning techniques to binary data is the lack of a large corpus of labeled training data. One solution to the lack of real-world data is to create synthetic data from real data through augmentation. In this paper, we demonstrate data augmentation techniques suitable for source code and compiled binary data. By augmenting existing data with semantically-similar sources, training set size is increased, and machine learning models better generalize to unseen data.