The goal of the study is to detect fake images. An autoencoder is trained adversarially to obtain a latent rich representation of real images. The latent representation is then used to reconstruct an image based on the given input image. We propose a comparative study to test if reconstructed features differ strongly for real and fake images leading to enhanced classification performance. We further argue that adversarially training a classifier (DCGAN) generalizes better on unseen data as compared to CNN based architectures.