Towards Effective Performance Fuzzing

Abstract

Fuzzing is an automated testing technique that utilizes injection of random inputs in a target program to help uncover vulnerabilities. Performance fuzzing extends the classic fuzzing approach and generates inputs that trigger poor performance. During our evaluation of performance fuzzing tools, we have identified certain conventionally used assumptions that do not always hold true. Our research (re)evaluates PERFFUZZ [1] in order to identify the limitations of current techniques, and guide the direction of future work for improvements to performance fuzzing. Our experimental results highlight two specific limitations. Firstly, we identify the assumption that the length of execution paths correlate to program performance is not always the case, and thus cannot reflect the quality of test cases generated by performance fuzzing. Secondly, the default testing parameters by the fuzzing process (timeouts and size limits) overly confine the input search space. Based on these observations, we suggest further investigation on performance fuzzing guidance, as well as controlled fuzzing and testing parameters.