The secure transmission in reconfigurable intelligent surface-assisted wireless communication system is investigated. Existing studies mainly focus on the ideal assumption that the eavesdropper's CSI is known, which is usually difficult to obtain in practice. Therefore, the secure transmission in reconfigurable intelligent surface-aided wireless communication system with unknown eavesdropper"s CSI is investigated. Firstly, the transmission power of the communication signal is minimized under the constraint of satisfying the quality of service of legitimate user through joint active/passive beamforming of beamforming vector and phase shift matrix, and then the residual power is allocated to artificial noise to jam potential eavesdropper. Deep learning-aided manifold optimization method is proposed to address the power allocation. This method combines the Riemannian gradient descent model with deep learning method. The adaptive learning ability of neural network is leveraged to dynamically learn the step size of the Riemannian gradient descent. Experimental results show that compared to existing optimization algorithms, the proposed method reduces the computational complexity by at least one order of magnitude while achieving almost identical secrecy rate.