Abstract
Coverage-guided graybox fuzzing is one of the most popular and effective techniques for discovering vulnerabilities due to its nature of high speed and scalability. However, the existing techniques generally focus on code coverage but not on vulnerable code. These techniques aim to cover as many paths as possible rather than to explore paths that are more likely to be vulnerable. When selecting the seeds to test, the existing fuzzers usually treat all seed inputs equally, ignoring the fact that paths exercised by different seed inputs are not equally vulnerable. This results in wasting time testing uninteresting paths rather than vulnerable paths, thus reducing the efficiency of vulnerability detection. In this paper, we present a solution, NeuFuzz, using the deep neural network to guide intelligent seed selection during graybox fuzzing to alleviate the aforementioned limitation. In particular, the deep neural network is used to learn the hidden vulnerability pattern from a large number of vulnerable and clean program paths to train a prediction model to classify whether paths are vulnerable. The fuzzer then prioritizes seed inputs that are capable of covering the likely to be vulnerable paths and assigns more mutation energy (i.e., the number of inputs to be generated) to these seeds. We implemented a prototype of NeuFuzz based on an existing fuzzer PTfuzz and evaluated it on two different test suites: LAVA-M and nine real-world applications. The experimental results showed that NeuFuzz can find more vulnerabilities than the existing fuzzers in less time. We have found 28 new security bugs in these applications, 21 of which have been assigned as CVE IDs.
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have