Publication Date
Spring 2026
Degree Type
Master's Project
Degree Name
Master of Science in Computer Science (MSCS)
Department
Computer Science
First Advisor
Fabio Di Troia
Second Advisor
Robert Chun
Third Advisor
Genya Ishigaki
Keywords
Opcodes, Diffusion, NLP, Graphs, GCN, GraphSAGE, GANs, WGAN, Word2Vec, FastText, Graph2Vec, Node2Vec, Embedding
Abstract
Malware has grown increasingly complex, and machine learning has become crucial for detection systems. However, their effectiveness is limited by the scarcity and diverse malware samples. Sample imbalance can lead to biased classifiers that underrepresent minority families and degrade performance. This project examines whether conditional generative models can produce synthetic opcode embeddings that can preserve the structural and semantic properties of real malware. Opcode sequences are represented with graph based and NLP based embeddings, which are used as inputs for a conditional WGAN-GP and conditional Diffusion model. The generated embeddings are evaluated through cross family injection, train real test synthetic, train synthetic test real, binary classification and class imbalance experiments. The results show that graph based embeddings consistently outperform NLP based ones, with the conditional Diffusion model producing the highest quality samples.
Recommended Citation
Macoco, Jonathan, "Evaluating Opcode based NLP and Graph Representations for Synthetic Malware" (2026). Master's Projects. 1815.
DOI: https://doi.org/10.31979/etd.zf9z-g59h
https://scholarworks.sjsu.edu/etd_projects/1815
