Neural Network-based Methodologies for Securing Cryptographic Code

Many studies show that manual code generation is error-prone and results in vulnerabilities. Vulnerability fixing has been shown as the most time-consuming process among multiple steps of code repair. To help developers repair these security vulnerabilities, my dissertation aims to develop an automatic or semi-automatic secure code generation system with neural network based approaches. Trained with huge amounts of good-quality code, I expect the neural network to learn the secure usage and produce the correct code suggestions.

Despite the great success of neural networks, the vision of comprehending and generating programming languages through neural networks has not been fully realized. There are many fundamental questions that need to be answered. These questions include 1) what are the accuracy impacts of the various choices in code embedding? 2) How to address the accuracy challenges caused by the programming language specific properties in the task of secure code suggestion? My dissertation work answers the two questions with a systematical measurement study and specialized neural network designs. My experiments show that program analysis is a necessary preprocessing step to guide the code embedding – resulting in a 36.1% accuracy improvement. Furthermore, I identify two previously unreported deficiencies in the cryptographic API suggestion task. To close the gap, I invent a highly accurate API method suggestion solution, referred to as Multi-HyLSTM, with specialized neural network designs to recognize unique programming language characteristics. My work points out the important differences between natural languages and programming languages, which pure data-driven learning approaches may not recognize.