ProvCreator: Synthesizing Graph Data with Text Attributes

Tianhao Wang, Simon Klancher, Kunal Mukherjee, Josh Wiedemeier, Feng Chen, Murat Kantarcioglu, Kangkook Jee

Keywords: Synthetic, Graph, Security, Intrusion Detection, Provenance

TL;DR: We develop a method for generating synthetic graphs with text attributes to support provenance-based intrusion detection systems.

Abstract:

In cybersecurity, system provenance graphs are a key primitive to support intrusion detection and program identification tasks. Recent movement towards using data-hungry graph learning models for security-critical applications has exposed significant limitations in existing provenance datasets. Imbalanced representation of programs induces bias and performance degradation in downstream models. Further, these models rely on rich numeric and textual node attributes to accurately encode program behaviors, limiting the ability of existing data augmentation techniques to address data imbalance in provenance graphs.

We present PROVCREATOR, a novel graph synthesis framework designed for feature-rich system provenance graphs. PROVCREATOR learns the joint distribution of node attributes and graph structures conditioned on program class labels, enabling targeted generation of realistic system provenance graphs to supplement underrepresented programs. Our evaluation shows that PROVCREATOR produces provenance graphs with higher structural fidelity, attribute fidelity, and downstream utility compared to those of previous graph synthesis methods.

Primary Area: other topics in machine learning (i.e., none of the above)

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 4736