Labelled Vulnerability Dataset on Android Source Code (LVDAndro) to Develop AI-Based Code Vulnerability Detection Models

Abstract

Ensuring the security of Android applications is a vital and intricate aspect requiring careful consideration during development. Unfortunately, many apps are published without sufficient security measures, possibly due to a lack of early vulnerability identification. One possible solution is to employ machine learning models trained on a labelled dataset, but currently, available datasets are suboptimal. This study creates a sequence of datasets of Android source code vulnerabilities, named LVDAndro, labelled based on Common Weakness Enumeration (CWE). Three datasets were generated through app scanning by altering the number of apps and their sources. The LVDAndro, includes over 2,000,000 unique code samples, obtained by scanning over 15,000 apps. The AutoML technique was then applied to each dataset, as a proof of concept to evaluate the applicability of LVDAndro, in detecting vulnerable source code using machine learning. The AutoML model, trained on the dataset, achieved accuracy of 94% and F1-Score of 0.94 in binary classification, and accuracy of 94% and F1-Score of 0.93 in CWE-based multi-class classification. The LVDAndro dataset is publicly available, and continues to expand as more apps are scanned and added to the dataset regularly. The LVDAndro GitHub Repository also includes the source code for dataset generation, and model training.