Adversarial samples are created by adding imperceptible perturbations to the original image that cause the image recognition model to make incorrect judgments. Unlike image adversarial samples, malware adversarial samples not only need to evade detection by the model but also must guarantee the executability and maintain the original behavior of the samples. With the rapid growth and widespread use of the Internet, the proliferation of malware has surged, particularly for Portable Executable (PE) files on the Windows platform. Existing malware adversarial sample generation methods can hardly meet the needs of detection engines in Artificial Intelligence (AI) enabled scenarios due to the problems of having non-mainstream objects, low bypass rate due to being trapped in local optimal solutions during the adversarial sample generation process, slow convergence speed during the generation process, and inability to guarantee the original behaviours. To solve these problems, we propose a new genetic algorithm-based method, Mal-POBM, to generate malware adversarial samples. The method improves the convergence speed by optimising the population, and jumps out the local optimal solution by bidirectional mutation. It ensures a high bypass rate of the adversarial samples while guaranteeing that the adversarial samples still have original behaviours. Meanwhile, we experimentally demonstrate the effectiveness of the method.