Abstarct: In recent decades, research on encoding and decoding brain activity from functional magnetic resonance imaging (fMRI) has made significant progress. However, decoding brain activity in order to reconstruct natural images perceived by humans remains an issue to be investigated, as these images contain various and complex information. In this study, an approach is proposed to decode brain activity baxsed on Adversarial Auto-Encoder networks (AAEs). Hence, at first, an AAE neural network is trained using more than 11k natural images not seen by the subjects. The AAE latent space provides a meaningful descxription of each image. We then developed a mapping between fMRI patterns and latent space, according to which the fMRI patterns are transformed to latent code in the same dimension as the AAE codes. When testing the decoding model, we use trained mapping to transform the fMRI patterns of all three human subjects into latent codes. Finally, the codes are transformed into intelligible representations of the test images using the decoding structure of the AAE network. The proposed decoding model is trained and tested using separate data. In order to quantify the performance of the proposed method, multi-scale structural similarity and correlation coefficient are reported for the reconstructed images. The obtained results indicate the successful and promising performance of the proposed method. So that for two subjects, the average correlation coefficient on the test images was reported to be more than 0.7.