Many wave and diffusion processes can be modeled with the Helmholtz equation, essential in wavefield computation. The unknown parameter identification is an efficient way to help researchers to understand the governing physics of a process. Classical methods for inverse problems require solving the forward equation many times, which leads to expensive computational costs as the model size increases. Recently, physics-informed neural networks (PINNs) have shown good performance in solving inverse problems due to their strong ability to represent PDEs and observed data. Benefits from the ability of neural networks to fit the observed data, there is no need to calculate the forward problem many times if we used classical methods. In this work, we identify the unknown space-varying parameter of wavenumber in the Helmholtz equation using physics-informed neural networks (PINNs). Through experiments, we also demonstrate the robustness of our method in handling high-noisy (up to 10%) data.