The Cost of Trust in Machine Learning - Privacy, Robustness, Unlearning, and their Interactions

Aug 5, 2026

As machine learning systems evolve from statistical tools to core societal infrastructure, their trustworthiness becomes a central scientific problem. The work shifts focus from maximizing accuracy to establishing formal guarantees for privacy, robustness, and unlearning. It studies these properties under a unified framework of quantifiable costs across utility, system assumptions, and computational resources. Results reveal both antagonistic trade-offs and synergistic effects, while also clarifying why unlearning is distinct from privacy yet can be achieved at lower utility cost, supporting a principled science of trustworthy learning.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages404

Reading Time16 h 58 min

AccessFree to download

Introduction and Background

The Interaction of Trust Guarantees

Research Questions

Thesis n° 11 375 The Cost of Trust in Machine Learning: Privacy, Robustness, Unlearning, and their Interactions Presented on 3rd November 2025 School of Computer and Communication Sciences Distributed Computing Laboratory Doctoral program in Computer and Communication Sciences for the award of the degree of Docteur ès Sciences (PhD) by Youssef ALLOUAH Accepted on the jury’s recommendation Prof. E. Telatar, jury president Prof. R. Guerraoui, thesis director Prof. A. Smith, examiner Prof. F. Bach, examiner Prof. G. Kamath, examiner Prof. S. Bengio, examiner 2025 Abstract As machine learning systems move from statistical tools to core societal infrastructure, their trustworthiness has become a primary scientific challenge. This requires a foundational shift from maximizing accuracy to providing formal guarantees on three critical properties: privacy, to protect the confidentiality of user data; robustness, to ensure integrity against malicious participants and data; and unlearning, to provide meaningful user control. This dissertation analyzes these pillars through a unified lens of quantifiable costs—in model performance (utility), system requirements (assumptions), and computation (resources) . Our analysis uncovers deep interactions between these trust guarantees, revealing them to be both antagonistic, as when privacy mechanisms hinder robustness, and synergistic, as when robust training provides a foundation for efficient unlearning. It also highlights fundamental separations, showing that unlearning, while related to privacy, is a distinct goal that can be achieved at a significantly lower utility cost. This work therefore provides a foundational map of these interactions, contributing towards a principled science of trustworthy machine learning. Keywords: machine learning, differential privacy, robust statistics, optimization, right-to-beforgotten, federated learning, distributed learning, collaborative learning Résumé À mesure que les systèmes d’apprentissage automatique passent du statut d’outils statistiques à celui d’infrastructures fondamentales de la société, leur fiabilité est devenue un défi scientifique majeur. Ceci exige un changement de paradigme, passant de la maximisation de la précision à l’obtention de garanties formelles sur trois propriétés critiques : la confidentialité, pour protéger la confidentialité des données des utilisateurs ; la robustesse, pour assurer l’intégrité contre les participants et les données malveillantes ; et le désapprentissage, pour offrir uncontrôle significatif aux utilisateurs. Cette thèse analyse ces piliers à travers le prisme unifié decoûts quantifiables, en performance du modèle (utilité), en exigences du système (hypothèses), et en calcul (ressources) . Notre analyse révèle de profondes interactions entre ces garanties de confiance, montrant qu’elles peuvent être à la fois antagonistes, comme lorsque les mécanismes de confidentialité entravent la robustesse, et synergiques, comme lorsque l’entraînement robuste sert de fondation à un désapprentissage efficace. Elle met également en évidence des séparations fondamentales, montrant que le désapprentissage, bien que lié à la confidentialité, est un objectif distinct qui peut être atteint à un coût nettement inférieur. Ce travail fournit ainsi une cartographie fondatrice de ces interactions, contribuant à une science rigoureuse del’apprentissage automatique fiable. Mots clefs : apprentissage automatique, confidentialité différentielle, statistiques robustes, optimisation, droit à l’oubli, apprentissage fédéré, apprentissage distribué, apprentissage collaboratif Acknowledgements I am grateful to all people and institutions who have contributed to or supported this thesis. I apologize for any unintentional omissions in naming them next. I thank Rachid Guerraoui for his generous support and career advice. He has provided me with a comfortable environment to freely develop my research tastes and

FAQs

What three trust properties does the dissertation focus on?

Privacy, robustness, and unlearning. Privacy protects user data confidentiality, robustness ensures integrity against malicious participants and data, and unlearning provides meaningful user control.

How does the dissertation evaluate privacy, robustness, and unlearning?

Through a unified lens of quantifiable costs across model performance (utility), system requirements (assumptions), and computation (resources).

What interactions between the trust guarantees are identified?

The work shows antagonistic interactions (e.g., privacy mechanisms hindering robustness) and synergistic interactions (e.g., robust training enabling efficient unlearning), and it also highlights that unlearning is a distinct goal with lower utility cost.

Aug 5, 2026

Upload by Valentina