Differentially-Private Federated Learning with Long-Term Budget Constraints Using Online Lagrangian Descent

This paper addresses the problem of time-varying data distribution in a fully decentralized federated learning setting with budget constraints. Most existing work cover only fixed data distribution in the centralized setting, which is not applicable when the data becomes time-varying, such as in realtime traffic monitoring. More so, a lot of existing work do not address budget constraint problem common in practical federated learning settings. To address these problems, we propose an online Lagrangian descent algorithm. To provide privacy to the local model updates of the clients, local differential privacy is introduced. We show that our algorithm incurs the best regret bound when compared to other similar algorithms, while satisfying the budget constraints in the long term.