Industrial Internet security is crucial for strengthening the manufacturing and network sectors of China. Deep learning, owing to its strong expression ability, good adaptability, and high portability, can support the establishment of an intelligent and autonomous industrial Internet security system and method. Therefore, it is of great value to promote the integrated innovation of deep learning and industrial Internet security. In this study, we analyze the development demand for industrial Internet security from the perspective of macro industrial environment, security technology, and deep learning system, and summarize the application status of deep learning to industrial Internet security in terms of device, control, network, application, and data layers. The security challenges faced by deep learning application to industrial Internet primarily lie in model training and prediction. Furthermore, we identify key research directions including interpretability of deep neural networks, cost control of sample collection and calculation, imbalance of sample sets, reliability of model results, and tradeoff between availability and security. Finally, some suggestions are proposed: A dynamic defense system in depth should be established in terms of overall security strategy; an application-driven and frontier exploration integrated method should be adopted to achieve breakthroughs regarding key technologies; and resources input should be raised for such interdisciplinary fields to establish an industry–university–research institute joint research ecosystem.