Optical coherence tomography (OCT) is being widely applied in clinical studies to investigate insight into the retina under the retinal pigment epithelium. Optical coherence tomography angiography (OCTA) is one of the functional extensions of OCT, for visualizing retinal circulation. Due to obstruction of light propagation, such as vitreous floaters or pupil boundaries, OCTA remains challenged by shadow artifacts that can disrupt volumetric data. Detecting and removing these shadow artifacts are crucial when quantifying indicators of retinal disease progression. We simplified an optical attenuation model of shadow formation in OCTA to a linear illumination transformation. And learn its parameters using an adversarial neural network. Our framework also consists of a sub-network for shadows automatic detection. We experimented our method on 28 OCTA images of normal eyes and compared the non-perfusion area (NPA), an index to measure retinal vascularity. The results showed that the NPA adjusted to a reasonable range after image processing using our method. Furthermore, we tested 150 OCTA images of synthesis artifacts, and the mean absolute error(MAE) values reached 0.83 after shadow removal.