Evaluation of the trade-off between performance and communication costs in federated learning scenario

Background and Objective: In traditional Machine Learning (ML) approaches, the data are collected and stored by a single node and subsequently used for training and testing. However, the acquisition and management of a large amount of data in some domains, as for example healthcare, can be problematic on account of the adoption of centralized architectures which entail certain security and privacy risks. Federated Learning (FL) has recently emerged as a technological solution to address such questions, even though communication efficiency may be a significant issue. Methods: This paper presents a novel learning strategy aimed at reducing the total number of parameters shared during the FL process and therefore, at evaluating a trade-off between the requirement to bring down communication costs and the need to guarantee the highest classification performance. Results: The results demonstrate the goodness of the solution in comparison with the traditional FedAvg algorithm since the accuracy of the proposed approach shows values ranging from 89.25% to 96.6% and, and in addition, the reduction of the communication overheads shows improvements ranging from 95.64% to 6%. Conclusion: The analysis of the proposed approach shows promising results in terms of performance and communication costs, especially in relation to the total amount of moved data since the challenge addressed by the paper concerns communication efficiency during the training process.