MARGIN: Margin-Aware Regularized Geometry for Imbalanced Vulnerability Detection

2026-07-09 · Source: cs.SE updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Cybersecurity & Data Privacy, Software Development & Engineering · Depth: Expert, extended

Summary

MARGIN (Margin-Aware Regularized Geometry for Imbalanced Vulnerability DetectioN) is a novel metric-based framework designed to address the severe frequency and difficulty imbalances prevalent in real-world software vulnerability datasets. The framework reinterprets these challenges from an embedding geometry perspective, observing that imbalances induce geometric distortions in hyperspherical representation space. MARGIN learns discriminative vulnerability representations through adaptive-margin metric learning and hyperspherical prototype modeling, dynamically adjusting geometric regularization based on von Mises–Fisher concentration. This aligns embedding distributions with Voronoi cells, reducing distortion and stabilizing decision boundaries. Extensive experiments on public datasets like BigVul, MegaVul, and ReposVul demonstrate that MARGIN consistently outperforms strong baselines, achieving notable improvements in classification and detection, particularly on challenging, imbalanced datasets, while also enhancing robustness, interpretability, and generalization.

Key takeaway

For AI Security Engineers building vulnerability detection systems, MARGIN offers a robust solution for highly imbalanced datasets. Its adaptive geometric regularization improves both binary detection and fine-grained CWE classification, especially for rare vulnerabilities. Consider integrating MARGIN to enhance model interpretability and generalization, ensuring more reliable security analysis and reducing false positives/negatives in critical scenarios.

Key insights

MARGIN geometrically regularizes hyperspherical embeddings with adaptive margins and prototype modeling to address imbalanced vulnerability detection.

Principles

• Imbalance distorts embedding geometry in hyperspherical space.
• Align vMF apex angle with Voronoi apex angle to reduce distortion.
• Adaptive margins and scaling stabilize decision boundaries.

Method

MARGIN uses CodeT5 for embeddings, estimates vMF concentration (kappa), and dynamically adjusts angular margins and logit scaling to align class distributions with Voronoi cells, improving separability.

In practice

• Utilize CodeT5 as a robust backbone for code embeddings.
• Estimate vMF kappa for class-wise concentration.
• Infer using geometric median prototypes for robustness.

Topics

• Vulnerability Detection
• Imbalanced Learning
• Metric Learning
• Hyperspherical Embeddings
• CodeT5
• Software Security

Code references

• stratum-dev/MARGIN

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, AI Security Engineer

Related on AIssential

• Sparsity-Inducing Divergence Losses for Biometric Verification — Q-Margin is a novel α-divergence loss that uses probabilistic margins for sparse, discriminative biometric embeddings.
• Tight Bounds for Learning Polyhedra with a Margin — A new PAC learning algorithm for polyhedra with a margin significantly improves efficiency and broadens applicability.
• Sinkhorn-CPD: Robust point cloud registration via unbalanced entropic optimal transport — Sinkhorn-CPD uses unbalanced entropic optimal transport with dual KL penalties for robust point cloud registration, automatically handling outliers and partial overlap.
• Geometry-Aware Bayesian Quantification via Compositional Data Analysis — A geometry-aware KDE model improves multiclass quantification by respecting the probability simplex.
• Lightweight Geometric Adaptation for Training Physics-Informed Neural Networks — Curvature-aware optimization using secant information significantly improves PINN training stability and accuracy.
• Conformal Margin Risk Minimization: An Envelope Framework for Robust Learning under Label Noise — CMRM improves noisy label learning by using a quantile-calibrated regularization term without privileged knowledge.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.SE updates on arXiv.org.