Directed grey-box fuzzing is a security testing technique aimed at discovering vulnerabilities in specific locations within a software system. In certain scenarios such as patch testing and error reproduction, directed grey-box fuzzing has demonstrated excellent performance. However, tools like AFLGO, representing directed grey-box fuzzing, typically follow specific distributions to select mutation operators, leading to significant wastage of computational resources by generating numerous invalid and redundant samples. This significantly reduces the efficiency of vulnerability detection in the tested software. This paper proposes a tool named Evfuzz, which is based on the Upper Confidence Bound (UCB) algorithm strategy to guide the seed mutation process in directed grey-box fuzzing. This approach aims to reduce randomness and blindness, achieving a more flexible and guided mutation. Experimental results demonstrate that, compared to the original directed grey-box fuzzer AFLGO, Evfuzz can reproduce crashes more quickly.