Behavioral Provenance Detection of Malicious Python Packages using Graph Neural Networks

Abstract

The increasing reliance on third-party packages from repositories such as Python Package Index (PyPI) and Node Package Manager (NPM) has introduced critical vulnerabilities in software supply chains. Traditional security approaches, including signature-based detection and trust evaluation systems, demonstrate limited effectiveness against sophisticated supply chain attacks such as dependency confusion, typo squatting, and account takeover. The study employs dynamic analysis to capture runtime behaviors during package installation, constructing provenance graphs that represent system interactions as nodes (processes, files, network connections) and edges (operations). A specialized Graph Neural Network architecture processes these graphs to distinguish malicious patterns from legitimate package behaviors. Preliminary results demonstrate detection accuracy exceeding 92% with false positive rates below 8%, significantly outperforming existing static analysis tools. The system provides interpretable detection results through attention mechanisms that highlight suspicious behavioral patterns, enabling security teams to understand and respond to threats effectively.

DOI: https://doi.org/10.5281/zenodo.21553162