Analyzing the Interplay Between Privacy, Explainability, and Fairness in Responsible Artificial Intelligence

You are cordially invited to attend the PhD Dissertation Defence of Fatima Ezzeddine on Friday 5 June 2026 at 09:00 in room D0.02.

Abstract:
Machine learning (ML) models are increasingly adopted in domains where decisions have significant social and economic impact, such as education, recruitment, and finance. As their use expands, the responsible deployment of such models has become a central concern. Beyond achieving strong predictive performance, deployed models should be audited and understood by humans, avoid reproducing or amplifying unfair biases, and safeguard the sensitive information contained in their training data. Recent regulatory frameworks, such as the European General Data Protection Regulation and the Artificial Intelligence (AI) Act, further reinforce these demands by mandating transparency, accountability, and safeguards for privacy and fairness in automated decision-making. As a result, data collection, model training, and deployment must satisfy not only technical performance criteria but also legal requirements, such as privacy, explainability, and fairness, for Responsible AI. Although substantial effort has been devoted to addressing each desiderata separately, achieving them simultaneously is considerably challenging. These requirements are often studied in isolation, despite their complex, interdependent interactions that can reinforce or undermine one another. For instance, enhancing transparency and explainability through explainable AI may expose sensitive information, while strong privacy guarantees can obscure explanations and hinder fairness auditing. Likewise, fairness interventions may alter model behavior in ways that affect explainability or increase privacy risks. In this thesis, we investigate how explainability interacts with other core Responsible AI principles to analyze the interplay among: (1) privacy and explainability, focusing on how explanations can be exploited for privacy attacks, such as membership inference and model extraction, and on corresponding mitigation strategies, including differential privacy and data distillation. We further investigate the synergies between collective privacy and explainability and identify associated open questions. (2) fairness and explainability, analyzing how fairness-aware modeling influences explanations and how the fairness of explanations themselves can be enforced. By characterizing the empirical trade-offs and identifying the conditions under which these desiderata can be jointly satisfied or necessarily conflict, this thesis contributes to the development of models that balance multiple Responsible AI requirements.

Dissertation Committee:
- Prof. Marc Langheinrich, Università della Svizzera italiana, Switzerland (Research Advisor)
- Prof. Silvia Giordano, SUPSI, Switzerland (Research co-Advisor)
- Dr. Omran Ayoub , SUPSI, Switzerland (Research co-Advisor)
- Prof. Gabriele Bavota, Università della Svizzera italiana, Switzerland (Internal Member)
- Prof. Paolo Tonella, Università della Svizzera italiana, Switzerland (Internal Member)
- Prof. Kévin Huguenin, University of Lausanne, Switzerland (External Member)
- Prof. Luca Longo, University College Cork, Ireland (External Member)
- Prof. Grégoire Montavon, Charité Berlin, Germany (External Member)
- Prof. Francesco Regazzoni, University of Amsterdam, Netherlands (External Member)