Federated Learning: Bridging Data Privacy and Model Accuracy on JointCloud

JointCloud is a novel collaborative cross-cloud architecture which is used to integrate and manager services of multiple clouds. Federated learning brings a promising intelligent scheme for JointCloud due to decentralization characteristic. However, data heterogeneity brings by JointCloud and causes the challenge of personalization and non-IID for traditional federated learning. To address the issue, we propose a novel Personalized Federated Learning method with Incentive mechanism (PFLI). We use reverse auction to incentive high-quality clients actively engage into training and customized the local models based on personalized aggregating situation. We evaluate our algorithm on two mainstream datasets and shows the advantage with three comparison baselines.