Multi-Argument Fuzzing by Reinforcement Learning

Fuzzing is a popular technique for testing software quality. This paper proposes a method to apply reinforcement learning to multi-argument fuzzing test. By tracking the rewards of each parameter to solve the MAB (Multi-Armed Bandit) problem of multi-argument selection and by using the average value of the total number of new paths and new crashes to determine the selection of exploration and exploitation stages, we can achieve a balance between these two stages to obtain the maximum reward.By this method, this paper implements a fuzzing test tool called MARL. In terms of edge coverage, the binary vim performed best, with MARL achieving 24.31 times higher coverage than MOPT and 14.27 times higher than AFL. Among the unique crashes, the binary nm-new performed best. Within the time limit, only MARL triggered 4 crashes, while the other fuzzers did not trigger any crashes.