Code reuse attack (CRA) is a powerful attack that reuses existing codes to hijack the program control flow. Control flow integrity (CFI) is one of the most popular mechanisms to prevent against CRAs. However, current CFI techniques are difficult to be deployed in real applications due to suffering several issues such as modifying binaries or compiler, extending instruction set architectures (ISA) and incurring unacceptable runtime overhead. To address these issues, we propose the first deep learning-based CFI technique, named DeepCheck, where the control flow graph (CFG) is split into chains for deep neural network (DNN) training. Then the integrity features of CFG can be learned by DNN to detect abnormal control flows. DeepCheck does not interrupt the application and hence incurs zero runtime overhead. Experimental results on Adobe Flash Player, Nginx, Proftpd and Firefox show that the average detection accuracy of DeepCheck is as high as 98.9%. In addition, 64 ROP exploits created by ROPGadget and Ropper are used to further test the effectiveness, which shows that the detection success rate reaches 100%.