Resilient Multimodal Fusion for Robust Vulnerability Exploitability Prediction Under Data Degradation

Abstract

Automated assessment of software vulnerability exploitability is essential for intelligent cyber defense, yet its effectiveness is often hindered by unstable, delayed, or degraded security data. Existing models typically rely on either unstructured descriptions or structured metadata, making them fragile to information gaps. We present SecureFusion‐X, a resilient multimodal framework designed for reliable prediction under adverse data conditions. The system utilizes decision‐level fusion with two specialized branches: TextGuard‐Lite, an efficient encoder based on a distilled MiniLM model, and TabShield‐Net, a TabTransformer‐based metadata encoder. A leakage‐free training protocol simulating modality perturbations supports robustness. Evaluation on 8549 curated NVD‐aligned vulnerability records shows that SecureFusion‐X improves F 1 from 0.9766 for the structured metadata baseline to 0.9798, while reducing the Brier score from 0.0202 to 0.0182 under clean‐test conditions. Under representative data degradation with 30% missing or noisy text, SecureFusion‐X remains stable, with less than 0.1% variation in fused performance, whereas the text‐only baseline drops by 9.5% under missing‐text conditions. SecureFusion‐X provides a computationally efficient and robustness‐oriented framework for vulnerability exploitability prioritization under controlled and temporally shifted evaluation settings.