From Detection to Prediction: Multi-Dimensional Embedding Similarity for Software Security

Abstract: Vulnerability detection remains a critical challenge in software security, particularly due to the limited availability of data needed to train deep learning models effectively.

In this talk, we present VulSim, a novel approach that leverages multi-dimensional neighbor embeddings to capture semantic, contextual and syntactic properties for improved function-level vulnerability detection. By integrating these diverse code properties, VulSim achieves strong accuracy and generalizability across datasets.

Building on this foundation, we will outline our ongoing research aimed at predicting future vulnerabilities by analyzing how vulnerabilities and their fixes co-evolve over time. This work includes constructing a dataset that captures real-world instances where security fixes inadvertently introduce new vulnerabilities. By mining these cases, we aim to uncover recurring patterns of vulnerability reintroduction and develop a proactive, pattern-driven prevention strategies.

Bio: Samiha Shimmi is a Ph.D. candidate in Computer Science at Northern Illinois University, advised by Dr. Mona Rahimi and a Research Aide Technical (Ph.D.) at Argonne.

Series: The CS Seminar Series is an event hosted by the Mathematics and Computer Science Division. The series invites a mix of speakers from within Argonne, as well as external speakers from the US and all over the world. Seminar talks span the breadth of Computer Science, with a particular focus on scientific computing, high performance computing and related topics.

See upcoming and previous presentations at CS Seminar Series.