Federated workload-aware quantized framework for secure learning in data-sensitive applications

Federated Learning (FL) emerged as a leading secure, distributed learning technology based on sharing insights instead of data. The privacy-ensuring capability of FL has enabled its extensive use in Data-Sensitive Applications like healthcare and finance. However, the transmitted insights are at risk of leakage as the security of the medium cannot be guaranteed and can lead to the inference of the user data. Quantization is sometimes used to change these transmitted values to provide security but at the cost of accuracy loss in global models. Coupled with client dropouts, this increases performance loss. In this paper, we propose a Federated Workload-Aware Framework with Linear Quantization (Fed-WALQ), which layers the quantization process with an active client-selection technique based on the sustainable workload of the clients. The framework minimizes the dropout rates and compensates for the loss due to quantization. Through numerical experiments compared against traditional FL and Quantization-enabled FL over multiple datasets, the Fed-WALQ shows improvements in security over the former and accuracy over the latter. The accuracy improvement varies with the complexities of the involved datasets, while a substantial drop in straggler node percentages is seen in all cases (up to 91.8% drop).