Suzzer: A Vulnerability-Guided Fuzzer Based on Deep Learning

Fuzzing is a simple and effective way to find software bugs. Most state-of-the-art fuzzers focus on improving code coverage to enhance the possibility of causing crashes. However, a software program oftentimes has only a fairly small portion that contains vulnerabilities, leading coverage-based fuzzers to work poorly most of the time. To address this challenge, we propose Suzzer, a vulnerability-guided fuzzer, to concentrate on testing code blocks that are more likely to contain bugs. Suzzer has a light-weight static analyzer to extract ACFG vector from target programs. In order to determine which code blocks are more vulnerable, Suzzer is equipped with prediction models which get the prior probability of each ACFG vector. The prediction models will guide Suzzer to generate test inputs with higher vulnerability scores, thus improving the efficiency of finding bugs. We evaluate Suzzer using two different datasets: artificial LAVA-M dataset and a set of real-world programs. The results demonstrate that in the best case of short-term fuzzing, Suzzer saved 64.5% of the time consumed to discover vulnerabilities compared to VUzzer.