Skip to content
Open access

Classifier-guided discrete diffusion for malware family-conditional graph generation

Oct 2026 · Journal of Computer Virology and Hacking Techniques · Vol 22 · 0 citations · 30 references

Abstract

We study whether discrete graph diffusion can be adapted to generate directed, weighted malware API call transition graphs in the low-data regime typical of dynamic malware analysis. Each execution trace is converted into a behavior graph whose nodes represent API types and whose directed edges represent transition count frequencies. We adapt a graph generative model, DiGress, to this setting by removing undirected symmetry constraints, modeling dense directed adjacency tensors with masking for variable graph sizes, and introducing an ordinal, marginal-aware diffusion kernel for edge-weight categories to respect the numeric structure of counts. We further enable malware-family-conditional generation using classifier guidance applied during sampling. Empirically, generated graphs preserve class-discriminative signal, but diverge strongly from real samples in edge-weight structure and multiple directed-graph diagnostics. In downstream tests, synthetic graphs are highly separable by malware family, yet a real-vs-synthetic detector achieves perfect accuracy, and classifiers trained on synthetic graphs transfer poorly to real data, indicating substantial distributional mismatch. These results highlight a failure mode of discrete diffusion over high-dimensional directed, weighted adjacency representations in a severely data-scarce setting.

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.