A Novel Federated Learning Scheme for Generative Adversarial Networks

Generative adversarial networks (GANs) have been advancing and gaining tremendous interests from both academia industry. With the development of wireless technologies, a huge amount data generated at network edge provides an unprecedented opportunity to develop GANs applications. However, due constraints such as bandwidth, privacy, legal issues, it is inappropriate collect send all cloud or servers for analysis, training, mining. Thus, deploying training becomes promising alternative solution. The instability introduced by non-independent identical (Non-IID) poses significant challenges GANs. To address these challenges, this paper presents novel federated learning framework GANs, namely, C ollaborated g xmlns:xlink="http://www.w3.org/1999/xlink">A me xmlns:xlink="http://www.w3.org/1999/xlink">P arallel Learning (CAP). CAP supports parallel models breaking isolated among generators that exists in previous distributed algorithms, achieving collaborative cloud, servers, devices. Then, further enhance ability CAP-GAN addressing Non-IID we propose Mix-Generator module (Mix-G) which divides generator into sharing layer personalizing layer. Mix-G extracts generic personalization features improves performance on extremely datasets. Experimental results analysis substantiate usefulness superiority our proposed scheme can achieve better scenarios compared with state-of-the-art algorithms.