该博客主要介绍了四个阶段：图形提取，句子嵌入，图像生成和分类。在图形提取阶段，通过静态分析函数源码得到程序依赖图。接着，将代码行视为句子并转化为固定长度向量。在图像生成步骤中，利用中心性分析重置向量，形成反映句子重要性的图像。最后，通过训练卷积神经网络模型对生成的图像进行漏洞检测。
Main Phases
Graph Extraction, Sentence Embedding, Image Generation, Classification
1. Graph Extraction
Main idea: normalize the source code of a function, and perform static analysis to extract the PDG ****of the function.
2. Sentence Embedding
Main idea: regard a line of code as a sentence and embed them into a fixed-length vector.
3. Image Generation
Main idea: After sentence embedding, apply centrality analysis to reset vector, which can reflect the difference between importance of sencences(or node of graph). use three centralities to correspond to the three channels of images and form the image.
4. Classification
Main idea: train a CNN model to detect the gerenated image to detect vulnerability.
1622
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
