As an opportunistic spectrum utilization technology, cognitive radio can greatly improve the spectrum utilization efficiency and alleviate the scarcity of spectrum resources. Spectrum sensing technique is key premise of realizing legitimate spectrum access in cognitive radio. In this paper, we propose to use a convolutional auto-encoder to solve the instability problem caused by complex environments in the traditional spectrum sensing process. The reconstruction error of deep learning model based on normal spectrum is an effective measure to judge whether the test signals are authorized or not. Moreover, the essential characterization capability of convolutional auto-encoder makes the metric well adapted to different environments and meet practical requirements. Finally, the effectiveness of the proposed method is verified by using a self-built broadcast dataset. Compared with state-of-the-art methods including PCA reconstruction, energy detection, and cyclostationary detection, the CAE based method shows better identification accuracy and robustness for unauthorized radio.