CMFuzz: context-aware adaptive mutation for fuzzers

Mutation-based fuzzing is a simple yet effective technique to discover bugs and security vulnerabilities in software. Given a set of well-formed initial seeds, mutation-based fuzzers continually generate interesting seeds by applying specific mutation strategy in order to maximize code coverage or the number of unique bugs explored at any point-in-time. However, existing fuzzers remain limited in the paths it could cover since it simply follows a uniform distribution to choose mutation operators. In this paper, we proposed a novel context-aware adaptive mutation scheme, namely CMFuzz, which utilizes a contextual bandit algorithm LinUCB to effectively choose optimal mutation operators for various seed files. To this end, CMFuzz dynamically extracts and encodes file characteristics, which allows mutation-based fuzzers to perform context-aware mutation. We apply this scheme on top of several state-of-the-art fuzzers, i.e., PTfuzz, AFL, and AFLFast, and implement CMFuzz-PT, CMFuzz-AFL, and CMFuzz-AFLFast, respectively. We conduct evaluation on 12 real-world open source applications and LAVA-M dataset against their counterparts. Extensive evaluations demonstrate that CMFuzz-based fuzzers achieve higher code coverage and find more crashes at a faster rate than their counterparts on most cases. Furthermore, we also utilize other mainstream bandit algorithms, e.g., Thompson Sample and epsilon-greedy, and implement Thompson-PT and Greedy-PT based on PTfuzz to examine the performance of proposed model. CMFuzz-PT significantly outperforms Thompson-PT especially in terms of unique crashes and paths, i.e., found 1.79× unique crashes and 1.29× unique paths on average. Compared to Greedy-PT, our approach still increases the amount of unique crashes and paths by 1.11× and 1.05×, respectively.