The outstanding performance of deep learning has prompted the rise of Machine Learning as a Service (MLaaS), which significantly reduces the difficulty for users to train and deploy models. For privacy and security considerations, most models in the MLaaS scenario only provide users with black-box access. However, previous works have shown that this defense mechanism still faces potential threats, such as model extraction attacks, which aim at stealing the function or parameters of a black-box victim model. To further study the vulnerability of publicly deployed models, we propose a novel model extraction attack named Generative-Based Adaptive Model Extraction (GAME), which augments query data adaptively in a sample limited scenario using auxiliary classifier GANs (AC-GAN). Compared with the previous work, our attack has the following advantages: adaptive data generation without original datasets, high fidelity, high accuracy, and high stability under different data distributions. According to extensive experiments, we observe that: (1) GAME poses a threat to victim models despite the model architectures and the training sets; (2) synthetic samples closed to decision boundary without deviating from the center of the target distribution can accelerate the extraction process; (3) compared to state-of-the-art work, GAME improves relative accuracy by 12% at much lower data and query costs without the reliance on domain relevance of proxy datasets.