Fisher-Guided Adaptive Fusion for Vulnerability Detection

A new method, Fisher-guided adaptive multimodal fusion, improves software vulnerability detection by addressing limitations of naive fusion of Natural Code Sequence (NCS) and Code Property Graph (CPG) representations. Empirical analysis shows pretrained models already encode structural information, causing overlap with CPG features from graph neural networks. Graph encoders are less effective than pretrained language models, so naive fusion can dilute discriminative cues. The proposed approach uses Fisher information to adaptively weight modalities, enhancing complementary signal extraction. The paper is available on arXiv.