文献类型：期刊文献

英文题名：FedGAN: Federated GAN for Few-shot Image Generation

作者：Zhang, Yanbing[1]; Zhang, Qian[1]; Yang, Mengping[1]; Xiao, Ting[1]; Wang, Zhe[1]

机构：[1] East China University of Science and Technology, Ecust, Department of Computer Science and Engineering, Shanghai, China

年份：2023

起止页码：1020

外文期刊名：2023 3rd International Conference on Electronic Information Engineering and Computer Science, EIECS 2023

收录：EI(收录号：20241015707108)

语种：英文

摘要：Few-shot image generation can generate new samples of the same class using a limited number of images. However, the progressive tightening of data privacy policies has made it difficult to obtain data for model training, which leads to the problem of poor quality of the generated samples. To address this problem, this paper proposes FedGAN, a few-shot generation model incorporating a federated framework. By combining the few-shot generation algorithm with a federated framework that protects data privacy, it applies model training based on data privacy protection and train generation adversarial networks by splicing Gaussian noise after image dimensionality reduction to achieve few-shot generation model training. Experiments results show that the designed few-shot generation method generates high-quality images on two datasets and compares with other non-federated methods achieving the best results on FID and LPIPS metrics. ? 2023 IEEE.