This thesis proposes novel approaches for the design and the implementation of trustworthy AI systems, considering both the unsupervised and supervised learning paradigms; it focuses on traditional ML models, less covered in the specialized literature compared to cutting-edge Deep Learning ones. As per unsupervised learning, a novel approach to execute the privacy-preserving C-Means and Fuzzy C-Means algorithms over decentralized data is presented, addressing both the horizontal and vertical data partitioning patterns. As per supervised learning, federated approaches are proposed for learning Takagi-Sugeno-Kang Fuzzy Rule-based Systems and Fuzzy Regression Trees, generally acknowledged as transparent or highly interpretable-by-design models for regression problems. This work gives an original contribution to the novel field of Fed-XAI (Federated Learning of eXplainable AI models) because it simultaneously addresses the requirements of privacy preservation and explainability, representing indeed a leap forward toward trustworthy AI. The contribution is not limited to algorithmic aspects: in fact, an extension of an existing open-source framework (named OpenFL) is also presented, aimed to support the implementation of Fed-XAI models. Such an extension, named OpenFL-XAI, provides a convenient method for creating AI applications balancing accuracy, privacy, and interpretability.
NEW APPROACHES TO LEARNING OF TRUSTWORTHY AI SYSTEMS IN CLUSTERING AND PREDICTION PROBLEMS
CORCUERA BÁRCENA, JOSÉ LUIS
2024
Abstract
This thesis proposes novel approaches for the design and the implementation of trustworthy AI systems, considering both the unsupervised and supervised learning paradigms; it focuses on traditional ML models, less covered in the specialized literature compared to cutting-edge Deep Learning ones. As per unsupervised learning, a novel approach to execute the privacy-preserving C-Means and Fuzzy C-Means algorithms over decentralized data is presented, addressing both the horizontal and vertical data partitioning patterns. As per supervised learning, federated approaches are proposed for learning Takagi-Sugeno-Kang Fuzzy Rule-based Systems and Fuzzy Regression Trees, generally acknowledged as transparent or highly interpretable-by-design models for regression problems. This work gives an original contribution to the novel field of Fed-XAI (Federated Learning of eXplainable AI models) because it simultaneously addresses the requirements of privacy preservation and explainability, representing indeed a leap forward toward trustworthy AI. The contribution is not limited to algorithmic aspects: in fact, an extension of an existing open-source framework (named OpenFL) is also presented, aimed to support the implementation of Fed-XAI models. Such an extension, named OpenFL-XAI, provides a convenient method for creating AI applications balancing accuracy, privacy, and interpretability.File | Dimensione | Formato | |
---|---|---|---|
Relazione_attivit_Corcuera.pdf
non disponibili
Dimensione
98.14 kB
Formato
Adobe PDF
|
98.14 kB | Adobe PDF | |
Thesis_PhD_jose_corcuera_final.pdf
embargo fino al 11/04/2027
Dimensione
2.44 MB
Formato
Adobe PDF
|
2.44 MB | Adobe PDF |
I documenti in UNITESI sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/20.500.14242/216085
URN:NBN:IT:UNIPI-216085