Towards Recovery of Conditional Vectors from Conditional Generative Adversarial Networks

TL;DR

This paper demonstrates that it is possible to recover latent and conditional vectors from conditional GANs, with promising results on both generated and real images, despite the challenges posed by deep neural network non-linearity.

Contribution

It introduces methods to recover latent and conditional vectors from conditional GANs and analyzes the differences in recovery performance between generated and real images.

Findings

Successful recovery of vectors from generated images

Identified performance gap between generated and real images

Promising qualitative and quantitative results

Abstract

A conditional Generative Adversarial Network allows for generating samples conditioned on certain external information. Being able to recover latent and conditional vectors from a condi- tional GAN can be potentially valuable in various applications, ranging from image manipulation for entertaining purposes to diagnosis of the neural networks for security purposes. In this work, we show that it is possible to recover both latent and conditional vectors from generated images given the generator of a conditional generative adversarial network. Such a recovery is not trivial due to the often multi-layered non-linearity of deep neural networks. Furthermore, the effect of such recovery applied on real natural images are investigated. We discovered that there exists a gap between the recovery performance on generated and real images, which we believe comes from the difference between generated data distribution and real data distribution. Experiments are conducted to evaluate the recovered conditional vectors and the reconstructed images from these recovered vectors quantitatively and qualitatively, showing promising results.