Personalized Online Federated Learning With Differential Privacy in Constrained Environments: A Meta-Learning Approach

Federated learning is a privacy-preserving technique that allows for collaboration among multiple clients to train a global model without data breaches. A notable challenge in the vanilla federated learning algorithm is the lack of personalization of the global model to any client, which is caused by data heterogeneity across the clients. Thus, there is no guarantee that the trained global model will perform optimally on any client's unseen tasks during test time. Although several works have addressed personalization in federated learning using various approaches, none of these works has been extended to the resource-constrained online learning setting with time-varying data distribution. Yet, this setting has a lot of practical applications, such as in wireless systems with time-varying channel conditions, as well as power and bandwidth limitations. Therefore, this paper proposes a Lagrangian optimization-based fully decentralized online federated learning algorithm that can provide model personalization with improved performance using meta-learning in the presence of long-term resource constraints. In addition, local differential privacy is included in the proposed algorithm to secure the models from reconstruction attacks. Simulation results compare the sublinearity of the regret bound with state-of-the-art algorithms.