Résumé : Narrowband radar micro-Doppler signatures are heavily used to identify and classify human activities. When the radar is operated in through-wall environments, the complex electromagnetic propagation phenomenology introduces considerable distortions in the micro-Doppler signatures through attenuation and multipath. The problem is particularly severe in inhomogeneous wall scenarios involving multiple wall layers, air gaps, or metal reinforcements. Through-wall radar data collection using simulations and measurements involves significant time and effort. In this paper, we propose an alternative method of synthesizing through-wall radar micro-Doppler signatures from their free space counterparts using the generative adversarial network (GAN). We train the GAN using radar micro-Doppler signatures generated from electromagnetic simulations. We generate the radar data for different human motions, along different orientations, and under diverse through-wall conditions. The synthetic radar micro-Dopplers generated from the neural networks are then evaluated for their realism using a denoising autoencoder, which shows an excellent realism score.