A Federated Generative Adversarial Network With SSIM-PSNR-Based Weight Aggregation for Consumer Electronics Waste

Federated learning-based GAN architectures are widely used to reduce resource intensiveness and ensure the privacy and security of individuals' data while generating synthetic images. Recently, several federated GAN architectures utilizing different aggregation mechanisms including FedAvg, FedSGD, MMD, etc., and heterogeneous model architectures were proposed to increase the generated images' quality. However, the presence of limited consumer electronics data and wide data acquisition diversity leads to model over-fitting resulting in compromised structural similarity and diversity of the generated images. Additionally, the iid distribution and differential learning on account of varied image diversity were not considered by previous works resulting in faulty aggregated weights which makes the architecture incapable of capturing diversity features in an edge device. To remedy the above concerns, this work proposes a federated GAN architecture with weighted aggregation mechanism based on a combination of SSIM and PSNR scores to ensure enhanced diversity in the generated dataset while increasing the structural similarity and overall image quality. The performance of the developed architecture is validated against three other existing state-of-the-art federated architectures in terms of SSIM and PSNR score. The experimental finding illustrates that the proposed architecture generates images with high quality, enhanced diversity, and better structural similarity.