Crowdsourcing in federated learning: a scalable approach to collaborative model training

Abstract

Mobile crowdsourcing, powered by advancing mobile technologies, offers efficient data management but poses privacy challenges. Federated Learning (FL) mitigates this by enabling joint model training while keeping data on-device [1]. However, sharing model parameters during training still risks data reconstruction by adversaries [2,3]. Differential Privacy (DP) is often used to protect FL models, yet its uniform deployment across heterogeneous clients can negatively impact performance, especially considering clients' heterogeneous privacy needs. The extended abstract explores two significant issues: (1). Trade-off between individualized privacy preservation and model utility, and (2). Constructing efficient perturbation mechanisms with minimal performance degradation.