Communication-Efficient Federated Learning for Resource-Constrained Edge/Iot Devices

Abstract

One of the most significant challenges for federated learning to be adopted in resource-limited Edge/IoT settings is communication overhead. The impact of Communication-Efficient FL Techniques (CEFLT) and Device Heterogeneity (DH) on the Communication-Efficient Federated Learning (CEFL) and Communication Efficiency (CE) as well as on the Edge/IoT System Performance (EISP) is investigated. The research design used was quantitative, explanatory and cross-sectional. The survey was conducted using a structured questionnaire, which was answered by 384 professionals, researchers, engineers and practitioners working in the fields of federated learning, edge computing and IoT technologies. Data was analysed using R Statistical Software. Descriptive statistics, reliability analysis, normality testing, correlation analysis, regression analysis, mediation analysis and moderation analysis were used to test the hypotheses.The results showed that CEFL is significantly influenced by CEFLT (β = 0.789, p < .001) and also DH significantly affects CEFL (β = 0.870, p < .001). CEFLT (β = 0.361, p < .001), DH (β = 0.402, p < .001), and CEFL (β = 0.130, p = .007) significantly improve CE. In addition, CE has a positive effect on EISP (β = 0.763, p < .001). The mediation analysis resulted in the confirmation of the proposed relationships, mediated by CEFL and CE; the moderation analysis showed that the relationships between communication and the other variables were strengthened by DH. 80% of the hypotheses were confirmed. This work builds upon Distributed Learning Theory and Edge Intelligence Theory, and shows that the communication efficiency is the main way on how communication-efficient federated learning improves the Edge/IoT system performance. The results offer both theoretical and practical contributions to the research and design of scalable, communication-efficient, and high-performance federated learning systems for heterogeneous Edge/IoT environments.