Improving Fitness Function for Language Fuzzing with PCFG Model

摘要

In this paper, we propose to use machine learning techniques to model the vagueness of bugs for language interpreters and develop a fitness function for the language fuzzing based on genetic programming. The basic idea is that bug-triggering scripts usually contain uncommon usages which are not likely used by programmers in daily developments. We capture the uncommonness by using the probabilistic context-free grammar model and the Markov model to compute the probabilities of scripts such that bug-triggering scripts will get lower probabilities and higher fitness values. We choose the ROC (Receiver Operating Characteristic) curves to evaluate the performance of fitness functions in identifying bug-triggering scripts from normal scripts. We use a large corpus of JavaScript scripts from Github and POC test cases of bug-reports from SpiderMonkey's bugzilla for evaluations. The ROC curves from the experiments show that our method can provide better ability to rank the bug triggering scripts in the top-K elements.