Considering the vulnerability of machine learning to adversarial attacks, the state-of-the-art malware detectors are designed with appropriate ensembles. However, most ensemble detectors are developed based on unreasonable and weak threat assumptions, which do not match the characteristics of real-world adversaries with powerful adaptive mixed capabilities. Such detectors will unavoidably suffer severe failures in practical deployment. To this end, we build two realistic powerful adversary models and propose NashAE as a robust malware detector based on a novel Game-theory-enabled ensemble adversarial training approach against them. Specifically, NashAE establishes a Minimax Game where the adversary and detector compete on opposing targets. By solving the Nash equilibrium of the game, NashAE can obtain the optimal ensemble adversarial training strategy under adversaries’ constantly adaptive attacks. Since the game has no closed-form solution, we further develop a simplified solution scheme based on Bayesian optimization to find the approximate Nash equilibrium of the game. We conduct comprehensive experiments with 10 baseline detection models on 2 malware datasets. Experimental results show that NashAE can achieve a stable detection rate of 58% higher than advanced methods after only 15 iterations against the most powerful adversaries.