Exploring Federated Learning for PerformancePrediction in Resource-Constrained Networks

Sammanfattning

Internet of things and its applicaitons is booming in many major industries. Equipping devices with a internet connection allows them to communicate with each other and make desirable automated decisions for us without human intervention. There are many problems that still needs to be faced with Internet of things technology, this thesis revolves around the problem assuring network performance. Typically nodes in wireless sensor networks have very limited recourses which makes it difficult to perform active performance measurments without overhead. The objective of this thesis is to add to previous research where network statistics are gathered so that machine learning algorithms can be applied to this data to predict the networkperformance indicator, round-trip time (RTT). This is done by exploring the benefits of applying federated learning compared to the traditional centralized learning in IoT networks. Then, experiments are conducted where this data is used to train and compare centralized learning models with federated learning models. When using all the data gathererd from all nodes in the testbed, the centralized learning model scored an f1-score of 0.77 while the federated learning model scored an f1-score of 0.59. A grouping strategy is devised where nodes with similar feature distributions are grouped together, this increased the federated learning model f1-score to 0.78 for one particiluar group while the centralized learning model scored the same f1-score of 0.77.Faculty of Scie nce and Tec hnolo gy, Up psala Unive rsity. Plac e of public ation U pps ala. Su p