Description
Points-to analysis is a subfield of static program analysis focused on identifying
which memory locations are accessed through pointers and references.
Due to the lack of runtime information, the problem is inherently undecidable
in a static context, resulting in reduced precision and an increased number of
false positives. This work introduces a novel approach that leverages learned
pointer relationships through a graph-based model. First, points-to information
is extracted through dynamic analysis, serving then as a ground truth. Subsequently,
a Graph Neural Network (GNN) is trained on an enriched dataset
represented as a Code Property Graph (CPG), which enables edge prediction
between code entities. The research aims to develop a learning-based approach
for pointer analysis, designed to achieve high precision through graph-based
modeling. The evaluation outcomes of the proposed model indicate a strong
potential for the application of GNNs in the domain of static analysis.
|
Thesis topic in co-operation with the Fraunhofer Institute for Applied and Integrated Security AISEC, Garching