A conditional Generative Adversarial Network allows for generating samplesconditioned on certain external information. Being able to recover latent andconditional vectors from a condi- tional GAN can be potentially valuable invarious applications, ranging from image manipulation for entertaining purposesto diagnosis of the neural networks for security purposes. In this work, weshow that it is possible to recover both latent and conditional vectors fromgenerated images given the generator of a conditional generative adversarialnetwork. Such a recovery is not trivial due to the often multi-layerednon-linearity of deep neural networks. Furthermore, the effect of such recoveryapplied on real natural images are investigated. We discovered that thereexists a gap between the recovery performance on generated and real images,which we believe comes from the difference between generated data distributionand real data distribution. Experiments are conducted to evaluate the recoveredconditional vectors and the reconstructed images from these recovered vectorsquantitatively and qualitatively, showing promising results.
展开▼