Abstract: With the rapid development of UAV as air relay and terminal, the issue of physical layer security has becoming a research hotspot in the recent. This article divides the communication power sent by the UAV into two parts that include power of confidential information and power of artificial noise, which can effectively prevent eavesdroppers eavesdropping illegally when UAV is transmitting confidential information. It also considers aerodynamic power consumption when UAV is flying. In order to maximize the average amount of information transferred at a fixed energy, this article optimizes the flight path and power distribution of the UAV. This scenario is modeled as a Markov model（MDP）, and uses the Curiosity-Driven Deep Q-learning Network（C-DQN） algorithm to achieve optimization. The result shows that the algorithm has a good convergent effect.