TSTFL-CGAN: Fuzzy Logic-Based Two-Stage Training Federated Learning Conditional Generative Adversarial Network Under Non-IID Data

With the proliferation of Internet of Things devices, federated Generative Adversarial Networks (GAN) have become an effective tool for handling data deficiencies in distributed architectures. However, in federated learning scenarios, due to the non-independently and identically distributed (non-IID) nature of data, the training of federated GAN faces multiple challenges. Existing works on federated GAN training still face limitations when confronted with highly heterogeneous data. To address these issues, this paper proposes a federated data augmentation and aggregation scheme based on Conditional GAN (CGAN), named TSTFL-CGAN. The method divides federated GAN training into two stages: pre-training and formal training. It uses fuzzy logic to handle uncertain data and allows the global server to flexibly allocate generators to supplement the missing samples on client. Additionally, we introduce a time factor to dynamically adjust each client’s model contribution to global aggregation based on the model training loss and data volume, which accelerates convergence and enhances the performance of the global model. Extensive experiments show that TSTFL-CGAN achieves convergence within 3 rounds of aggregation across 4 different data distribution scenarios, and it demonstrates an improvement of 10 to 30 in the evaluation of generated sample quality.