Deepfake technology generates highly realistic videos effortlessly, raising serious concerns about privacy violations, misinformation, and financial fraud. Detecting Deepfakes is the most effective solution to address these issues. While existing detection methods perform well on standard protocols, they struggle with real-world scenarios due to constantly emerging unknown forgery types. To enhance detection generalization, recent methods augment training images by synthesizing diverse forged faces (pseudo-fake faces) and identifying common forgery features. In this paper, we describe a new augmentation-based method to further improve the detection generalization. Our method leverages adversarial generative learning, which adaptively synthesizes effective pseudo-fake faces based on a generator network, a face synthesizer, a face reconstruction, and a discriminator. Extensive experiments on several public datasets demonstrate the efficacy of our method.