Fuzzing is an effective method to find software bugs and vulnerabilities. One of the most useful techniques is the coverage-guided fuzzing, whose key element is the tracing code coverage information. Existing coverage-guided fuzzers generally use the the number of basic blocks or edges explored to measure code coverage. Path-coverage can provide more accurate coverage information than basic block and edge coverage. However, the number of paths grows exponentially as the size of a program increases. It is almost impossible to trace all the paths of a real-world application. In this paper, we propose a fuzzing solution named PathAFL, which assists a fuzzer by path identification. It can effectively identify and utilize the important h-path, which is a new path but whose edges have all been touched previously. First, PathAFL only inserts one assembly instruction to AFL's original code to calculate the path hash, and uses a selective instrumentation strategy to reduce the tracing granularity of an execution path. Second, we design a fast filtering algorithm to choose higher weight paths from a large number of h-paths and add them to the seed queue. Third, both the seed selection algorithm and the power schedule are implemented based on the path weight. Finally we implemented PathAFL based on the popular fuzzer AFL and evaluated it on 10 well-fuzzed benchmark programs. In 24 hours, PathAFL explored 38% more paths and 9.3% more edges than AFL. Compared with CollAFL-x, the number is 25% and 5.9% correspondingly. Moreover, PathAFL found the more bugs on the LAVA-M dataset, even four unlisted bugs. The results show that PathAFL outperforms the previous fuzzers in terms of both code coverage and bug discovery. In well-tested programs, PathAFL found 8 new security bugs with 6 CVEs assigned.

No abstract