The paper revisits Neural Program Smoothing (NPS) for fuzzing, a machine learning-guided fuzzing method, and conducts an extensive evaluation against standard gray-box fuzzers. The authors found that the original performance claims for NPS fuzzers do not hold due to several limitations. They implemented Neuzz++, which addresses some of these limitations and improves performance. However, they found that standard gray-box fuzzers almost always surpass NPS-based fuzzers. The paper proposes new guidelines for benchmarking fuzzing based on machine learning and presents MLFuzz, a platform for the evaluation of ML-based fuzzers.
Publication date: 29 Sep 2023
Project Page: https://doi.org/10.1145/3611643.3616308
Paper: https://arxiv.org/pdf/2309.16618
