MGFL: Multi-granularity Federated Learning in Edge Computing Systems

As a promising machine learning framework in the big data era, federated learning (FL) allows multiple mobile devices to collaboratively train a model without transmitting raw data, thus has attracted widespread attention in both academia and industry. Considering that heterogeneous mobile devices with limited resources and data diversity are bound to impact the actual performance of some training nodes. However, conventional FL could not support collaborative training with multi-granularity neural networks. To this end, we propose multi-granularity federated learning (MGFL) that contains two mechanisms serving for same-granularity FL and cross-granularity FL. MGFL customizes a personalized model for each device by designing a divergence-based similarity measurement method in same-granularity FL. Further, it adjusts the empirical risk loss function to break the restriction of cross-granularity FL. Experimental evaluations demonstrate the positive guidance of the fine-granularity model to the coarse-granularity model, which significantly improves the performance of the coarse-granularity model. Besides, our method shows superiority on both independently identically distribution (IID) and non-IID data.