FedMVA: Enhancing software vulnerability assessment via federated multimodal learning

Qingyun Liu, Xiaolin Ju, Xiang Chen, Lina Gong

Abstract: Highlights•Proposes a federated multimodal vulnerability assessment method (FedMVA).•Integrates code structure, lexical features, and developer comments for comprehensive assessment.•Employs a weighted variance minimization loss to improve global-local model alignment.•Enhances robustness with dynamic learning rates and momentum-based client weighting.•Ablation studies demonstrate that FedMVA consistently outperforms existing baselines.