Generative Federated Learning for Smart Prediction and Recommendation Applications
Anwesha Mukherjee, Rajkumar Buyya
TL;DR
The paper tackles the challenge of privacy-preserving, low-latency heart health prediction under data scarcity and non-IID conditions. It introduces Generative Federated Learning (GFL), integrating GAN-based synthetic data generation with both centralized (CFL) and decentralized (DFL) federated learning, using GRU for sequential health signals. Empirical results on a heart health monitoring case show ~99% accuracy and substantial latency reductions (~73% vs cloud-only) when employing GAN-augmented FL across edge devices, with CFL and DFL achieving comparable performance but differing in topology and overhead. The work demonstrates the practicality of privacy-preserving, data-diverse, and fast health analytics at the edge, while highlighting future directions for lightweight data generation, federated offloading, and Bayesian methods to further improve robustness.
Abstract
This paper proposes a generative adversarial network and federated learning-based model to address various challenges of the smart prediction and recommendation applications, such as high response time, compromised data privacy, and data scarcity. The integration of the generative adversarial network and federated learning is referred to as Generative Federated Learning (GFL). As a case study of the proposed model, a heart health monitoring application is considered. The realistic synthetic datasets are generated using the generated adversarial network-based proposed algorithm for improving data diversity, data quality, and data augmentation, and remove the data scarcity and class imbalance issues. In this paper, we implement the centralized and decentralized federated learning approaches in an edge computing paradigm. In centralized federated learning, the edge nodes communicate with the central server to build the global and personalized local models in a collaborative manner. In the decentralized federated learning approach, the edge nodes communicate among themselves to exchange model updates for collaborative training. The comparative study shows that the proposed framework outperforms the existing heart health monitoring applications. The results show that using the proposed framework (i) the prediction accuracy is improved by 12% than the conventional framework, and (ii) the response time is reduced by 73% than the conventional cloud-only system.