基于深度学习的多模态漏洞检测方法

With the popularization of computer technology, code security issues have become a key factor restricting its development. This article proposes a multi-modal fusion vulnerability detection model based on deep learning. It combines a bidirectional recurrent neural network support vector machine (BiLSTM-SVM), a serial parallel convolutional bidirectional recurrent network model, an ensemble classification algorithm (CNN-BiLSTM-EC), and a hybrid model (MultiFusionNet), and establishes an accurate correspondence between source code and assembly code through a code alignment algorithm, fusing them into a hybrid code. Finally, through a multi-modal hybrid fusion strategy, it preserves the unique information of each feature and the combined features across features, providing the most comprehensive code representation. Experiments based on SARD data sets show that this method achieves an average performance of more than 94% for F1 core, which is significantly better than the single mode method. This study provides a new technological path for code security detection.