Facial aging and facial rejuvenation analyze a given face photograph to predict a future look or estimate a past look of the person. To achieve this, it is critical to preserve human identity and the corresponding aging progression and regression with high accuracy. However, existing methods cannot simultaneously handle these two objectives well. We propose a novel generative adversarial network based approach, named the Conditional Multi-Adversarial AutoEncoder with Ordinal Regression (CMAAE-OR). It utilizes an age estimation technique to control the aging accuracy and takes a high-level feature representation to preserve personalized identity. Specifically, the face is first mapped to a latent vector through a convolutional encoder. The latent vector is then projected onto the face manifold conditional on the age through a deconvolutional generator. The latent vector preserves personalized face features and the age controls facial aging and rejuvenation. A discriminator and an ordinal regression are imposed on the encoder and the generator in tandem, making the generated face images to be more photorealistic while simultaneously exhibiting desirable aging effects. Besides, a high-level feature representation is utilized to preserve personalized identity of the generated face. Experiments on two benchmark datasets demonstrate appealing performance of the proposed method over the state-of-the-art.