With the development of image acquisition technology, the volume of image data has surged, highlighting the contradiction between data publication and privacy protection. Generative Adversarial Networks (GANs) offer a solution to find a balance between the development of image data and privacy security. However, the unidirectional image generation of GANs fails to satisfy the reversible requirements of privacy-sensitive images. To address this limitation, this study proposes an image privacy protection method based on a dual-prior GAN with dynamic reversibility, called DR-DPGAN. This method uses StyleGAN2 latent space editing to make targeted modifications to image identity features, allowing modification and reconstruction of features. To achieve image privacy protection, a fake identity generator composed of two-layer multi-layer perceptrons is designed. By combining identity-related guidance information, it precisely controls the generation of fake features to avoid excessive or insufficient modification. Meanwhile, three-dimensional prior constraints are introduced to extract geometric feature vectors, maximizing the retention of original non-identity attribute features and ensuring the usability of images in downstream tasks. To ensure reversible image restoration, this paper converts the original identity attribute information into binary vectors through a binary encoding mapping network, generating reversible encrypted features to ensure precise restoration of original identity features. In addition, four loss functions are used jointly to optimize the network to balance the quality of the generation. To verify the reversibility and effectiveness of the proposed method, comprehensive experimental tests are conducted on two different datasets. The experimental results demonstrate the effectiveness of this method in image anonymization and reversible restoration.