Dynamic Resource-Aware Federated Framework for Secure and Sustainable Learning in Data Sensitive Applications

Federated Learning (FL) quickly gained popularity as a secure distributed learning technology based on insights sharing. The client-level privacy provided by FL has enabled its extensive integration in fields dealing with susceptible data such as E-commerce, the Internet of Medical Things, etc. However, the transmitted insights may still be subjected to eavesdropping since the security of the transmission medium cannot be guaranteed. Using complex techniques, this can lead to the inference of the client data. In many cases, quantization is applied over transmitted payload to hinder such attempts but results in drop of accuracy in the global models. Moreover, the capabilities of client nodes are not considered in most implementations which may lead to under or over-utilization of client resources and cause client dropouts due to exhaustion. When combined, the quantization and client dropouts cause a significant performance loss and make the network less stable. In this paper, we propose a Dynamic Resource-Aware Federated Framework for Secure and Sustainable Learning, which layers the quantization process with an active workload management technique based on the capabilities of the client nodes. The framework minimizes the dropout rates significantly while recovering the accuracy lost in quantization to some extent. Through experiments compared against standard FL and Quantized FL over multiple datasets, the DRAFFSS shows gains in security over the former and in accuracy over the latter while keeping client dropouts negligible.