Detection of Software Vulnerabilities using Machine Learning

早期发现软件漏洞是保证计算机系统安全的关键。虽然某些类型的安全漏洞可以自动识别，但大多数漏洞仍然是通过耗时的手动分析发现的。该项目旨在利用机器学习技术开发新的漏洞发现方法。为此，代码的结构化表示，如解析树和控制流图，被嵌入到向量空间中，并使用无监督学习算法进行语义分析和异常检测。这种分析应该能够自动识别典型的编程模式，并检测偏离这些模式的潜在漏洞。

The early detection of vulnerabilities in software is a key for securing computer systems. While some types of security flaws can be identified automatically, the majority of vulnerabilities is still discovered by time-consuming manual analysis. This project aims at developing novel methods for vulnerability discovery using machine learning techniques. To this end, structured representations of code, such as parse trees and control flow graphs, are embedded in vector spaces and analyzed using unsupervised learning algorithms for semantic analysis and anomaly detection. This analysis should enable identifying typical programming patterns automatically and detecting potential vulnerabilities as deviations from these patterns.