GeMuFuzz: Integrating Generative and Mutational Fuzzing with Deep Learning

Current grey-box protocol fuzzers may not work well with poor-quality initial seeds. That makes it difficult to cover diverse message types and protocol states defined in the protocol specification. To mitigate this issue, we propose GeMuFuzz, which integrates deep learning based seed generation into mutation-based grey-box fuzzing. Moreover, GeMuFuzz considers the high-dimensional information implied in seeds generated during fuzzing. We also evaluated the performance of GeMuFuzz by comparing with the baseline fuzzer AFLNET on 8 typical protocol implementations of ProFuzzBench. GeMuFuzz discovered 5.07% more paths and 6.19% more crashes, as well as 8.57% more states and 10.54% more state transitions than AFLNET. The experimental results highlight that GeMuFuzz could improve the effectiveness of fuzzing.