Abstract: In order to solve the problem that the discriminator loss value of traditional adversarial neural networks in GAN (Generative Adversarial Networks) model of Zero-sum Game is too high, the cost function of adversarial neural networks is improved.In the incomplete information of the Zero-sum Game, because the MNIST digital image information than what the generator and the discriminator of the game confrontation know each other is limited, to analyze the role of certain factors in the confrontation of the gaming process, the cost function particularly important in the Generative Adversarial Networks model of Zero-sum Game. Research on the selection of cost function and performance evaluation in the confrontation environment of the Zero-sum Game, and then realize the Zero-sum Game with greater advantages. Experimental results show that the cost function of the improved Generative Adversarial Networks is better than the traditional Generative Adversarial Networks loss function method.