User selection strategies for improved federated learning over wireless
Xu, Weijie (2023)
Xu, Weijie
2023
Master's Programme in Information Technology
Informaatioteknologian ja viestinnän tiedekunta - Faculty of Information Technology and Communication Sciences
This publication is copyrighted. You may download, display and print it for Your own personal use. Commercial use is prohibited.
Hyväksymispäivämäärä
2023-12-11
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:tuni-2023120710505
https://urn.fi/URN:NBN:fi:tuni-2023120710505
Tiivistelmä
Traditional Machine Learning (ML) methods often centralize data on a single server for processing and analysis, which poses significant risks in terms of data privacy. In contrast, Federated Learning (FL) offers an alternative that protects user privacy, allowing collaborative model training across multiple clients while keeping the data localized. One of the main challenges faced by FL is substantial communication overhead, especially in cases where clients have varying computation capacities and network conditions. This study examines the performance differences in terms of total latency and generalization capability among Random Scheduling (RS), Round Robin (RR), and Proportional Fair (PF) strategies, based on 10 simulation runs in Independently and Identically Distributed (IID) and Non-Independently and Identically Distributed (non-IID) data environments.
The simulation results show that RS and RR strategies are similar in terms of the average total latency, but RR strategy displays more variability in total latency. In contrast, the PF strategy significantly reduces communication latency. Increasing the number of clients chosen per iteration considerably improves the generalization capabilities of models using RS or RR. However, models using the PF strategy did not exhibit significant improvements in generalization when increasing the number of selected clients in scenarios with a large number of clients. Therefore, the PF strategy might be more efficient in environments with fewer users and greater variation in client performance compared to the RS and RR strategies. In scenarios with more users, RS and RR strategies may be preferable.
The simulation results show that RS and RR strategies are similar in terms of the average total latency, but RR strategy displays more variability in total latency. In contrast, the PF strategy significantly reduces communication latency. Increasing the number of clients chosen per iteration considerably improves the generalization capabilities of models using RS or RR. However, models using the PF strategy did not exhibit significant improvements in generalization when increasing the number of selected clients in scenarios with a large number of clients. Therefore, the PF strategy might be more efficient in environments with fewer users and greater variation in client performance compared to the RS and RR strategies. In scenarios with more users, RS and RR strategies may be preferable.