VulnScout: Imbalance-aware source code vulnerability detection using pretrained code models and Deep Neural Network

2026-06-13

PRE

Muhammad Farhat Ullah

Maseeh Ullah Khan

Ali Saeed

Sabeeh Ullah Khan

Muhammad Ishtiaq

Hasan E. Rezwan

Weiqiang Kong *

DOI：10.1016/j.infsof.2026.108228

Original

Original request for help

Save

Abstract

En 中文

• A lightweight model detects source code vulnerabilities with minimal resources. • Combines frozen transformer embeddings with adaptive focal loss for skewed data. • Handles long functions using overlapping chunking and mean pooling strategy. • Outperforms fine-tuned LLMs and GNNs on six public benchmarks. • Generalizes well to unseen codebases with high F1 and low false-positive rates.

VulnScout: Imbalance-aware source code vulnerability detection using pretrained code models and Deep Neural Network

Abstract

Journal

Organization

Cited Papers

Citing Papers