• federated learning
• explainable artificial intelligence
• machine learning
• fuzzy models
• tsk
• fuzzy regression tree
• openfl

We are now living through an Artificial Intelligence Golden Age driven by
three major forces: a massive, ever-growing, data availability, computational
power to process this data and algorithmic advancements in the Machine Learn-
ing field. Nevertheless, today’s AI is facing some major dragging forces too: the
awareness for data privacy concerns is increasing and there is the need for AI
algorithms to be explainable. To this extent, the work of this thesis consists in de-
signing, implementing and testing an AI system, fulfilling, at once, the following
requirements: explainability, data privacy and performance. Customized Feder-
ated Learning schemes for Fuzzy-Rule-Based Systems as eXplainable AI (XAI)
models are proposed, within the broader context of a distributed FED-XAI ap-
plication research project. After an application requirements analysis has been
carried through in order to derive an architecture design proposal, the effort was
focused on the model learning module. A feasibility study has been conducted
to evaluate existing works on i) Federated Learning frameworks, ii)XAI solutions
and iii)FED-XAI solutions combining both worlds together. As results of this
study, OpenFL 1 , an open-source Federated Learning framework by Intel, has been
chosen, reverse engineered and extended with XAI models support. A Takagi-
Sugeno-Kang Fuzzy Rule-Based system, customized to be learned in a federated
fashion while preserving its interpretability, has been chosen as XAI model and
ported within OpenFL. Subsequently, a centralized multi-way Fuzzy Regression
Tree has been explored as next XAI model. After an extensive experimental anal-
ysis on the tree depth parameter, the model has been successfully ported within
OpenFL, according to a federation scheme similar to the one adopted for TSK
System. A thorough experimental analysis has been performed, on ten bench-
mark regression datasets, to validate this work by highlighting the achievements
obtained and future directions to take.
This thesis has been partly developed under the framework of the H2020
project Hexa-X (Grant Agreement no. 101015956) founded by the European
Commission.