[Objective] By CycleGAN and improved generation loss through Wasserstein distance improving the stability and accuracy of the rumor detection model in cases of imbalanced and unpaired data samples. [Methods] The rumor discriminative model is enhanced by adversarial training between generator and discriminator. We use Wasserstein distance upgrade the cycle consistency loss and identity loss, and introduce both loss in generation training process to achieve controllability of the generated target, and improving the guidance effect of the generator while avoiding the problem of gradient explosion that may occur during adversarial network training. [Results]Our method achieves an accuracy of 0.8539 and an F1 value of 0.8376 on the unbalanced rumor dataset PHEME, and compared with the latest method, it has increased by 0.0068 and 0.018 respectively. [Limitations] The rumor detection model based on CycleGAN only has two generators and can only achieve the conversion of two categories of samples, so it just suitable for binary classification rumor detection models, and cannot be applied to the multi classification rumor detection tasks. [Conclusions] Our proposed model can effectively enhance the rumor detection ability of the rumor detection model in cases of imbalanced data.