RLF: Directed Fuzzing based on Deep Reinforcement Learning

Fuzzy testing is widely used in the field of vulnerability mining due to its ease of use and low false positive errors. However, traditional fuzzy tests suffer from low efficiency, blind test sample generation, and huge resource consumption. In order to solve these problems, we propose RLF‐‐a deep reinforcement learning-based directed fuzzing. RLF uses program instrumentation to obtain the execution path and calculates the distance between the path and the target. RLF will apply a deep reinforcement learning network that performs optimization to guide the selection of test samples basing on the information. Experimental results show that the method has better efficiency than other fuzzy testing tools in terms of vulnerability reproduction and significantly improves the orientation of fuzzy tests.