Website Fingerprinting (WF) attack is an attack on encrypted web traffic. The attacker recognizes different websites through analyzing the flow-based features extracted from encrypted traffic. Despite many defenses have been developed, most methods have the disadvantages of high overhead or poor defense effectiveness. Specifically, the newest WF attack based on deep neural network defeats those defenses by learning the defense strategy. In this paper, we proposed WF-GAN, a WF defense based on Generative Adversarial Networks (GANs). Our approach automatically generates adversarial traffic feature by adversarial learning combing with a WF classifier. The experimental result shows that WF-GAN achieve 90\% success rate with 15\% overhead on any fractions of the source websites set, which outperforms than previous defense. In addition, we proposed a new defense level, targeted defense, which does not support by previous defense. The result shows that the targeted defense success rate of WF-GAN is over 90\% when the target websites set is two times than the source website set.