A Novel Conditional Wasserstein Deep Convolutional Generative Adversarial Network
Arunava Roy, Dipankar Dasgupta · IEEE Transactions on Artificial Intelligence · 2023
Generative Adversarial Networks (GAN) and their several variants have not only been used for adversarial purposes but also used for extending the learning coverage of different AI/ML models. Most of these variants areunconditionaland do not have enough control over their outputs.ConditionalGANs (CGANs) have the ability to control their outputs byconditioningtheirgeneratoranddiscriminatorwith an auxiliary variable (such asclass labels, andtext descriptions). However, CGANs have several drawbacks such asunstable training,non-convergenceand multiplemode collapses like otherunconditionalbasic GANs (where thediscriminators areclassifiers). DCGANs, WGANs, and MMDGANs enforce significant improvements to stabilize the GAN training although have no control over their outputs. We developed a novelconditionalWasserstein GAN model, called CWGAN (a.k.aRD-GANnamed after the initials of the authors' surnames) that stabilizes GAN training by replacingrelatively unstableJS divergence with Wasserstein-1 distance while maintaining better control over its outputs. We have shown that the CWGAN can produce optimalgenerators anddiscriminators irrespective of the original and input noise data distributions. We presented a detailed formulation of CWGAN and highlighted its salient features along with proper justifications. We showed the CWGAN has a wide variety of adversarial applications including preparingfakeimages through a CWGAN-baseddeep generative hashing functionand generating highly accurate user mouse trajectories for fooling any underlying mouse dynamics authentications (MDAs). We conducted detailed experiments using well-known benchmark datasets in support of our claims.