FEDERATED LEARNING APLICADA EM DISPOSITIVOS COM TECNOLOGIA INTERNET OF THINGS
Coelho, Mariana França
MSc Thesis · 2022 · University of Coimbra
Supervisors: Pedro Manuel Henriques da Cunha Abreu, Tiago José dos Santos Martins da Cruz
Abstract
Federated Learning has gained momentum recently because it is highly scalable and can host different devices with heterogeneous data. It also allows collaboration between multiple entities without exposing data. Devices, which have Internet of Things technology, are increasingly present in the daily lives of humans, demonstrating the need for high information processing. From this, we focus on exploring and understanding the implementation of Federated Learning with different frameworks relative to the range of Internet of Things devices. The experiments were performed using the framework Flower to evaluate the parameters. With this, the performance of the PassiveAggressive, Perceptron, SVM, and Logistic Regression models were tested for all UNSW-NB15 (2015), CIC-IDS2017 (2017), TON_IoT (2019), WUSTL-IIOT-2021 (2021) datasets, which provide information of the presence of cyber attacks. We sought to explore the simulated environment in applications of a binary, and multiclass classifier, from Federated Learning techniques, addressing each of the steps and understanding the primary evaluation metric for decision making, f1-score. We run combinations between the presented models and dimensionality reduction techniques to calculate the metric. We present the results obtained from the simulations and the challenges encountered during the steps, concluding that in almost all tests, we obtained better results with the PassiveAggressive model, besides the dimensionality reduction from the importance of the characteristics. For each test performed, we reached an information gain among clients of at least 2% f1-score, denoting the effectiveness of the simulated environment. For future work, we suggested applying scenarios using unbalancing the number of samples distributed to all clients.