Optimizing Fairness in Federated Learning: Balancing Fairness and Performance under Budget Constraints ​

Machine Learning (ML) is now widely used in domains such as finance, healthcare, and criminal justice, where algorithmic decisions directly affect opportunities, rights, and access to resources. These systems can achieve remarkable accuracy, but models trained on biased data may also reproduce and amplify existing inequalities. Ensuring that ML is not only accurate but also fair has therefore become a pressing concern. At the same time, legal and ethical restrictions often prevent the centralization of sensitive data, which has led to the rise of Federated Learning (FL), a paradigm that enables collaborative training without sharing raw data. Although FL limits the need to centralize data by keeping records local, it raises distinctive challenges for fairness: clients typically hold non-IID data, meaning that their datasets differ in size, composition, or distribution, which can exacerbate disparities. Moreover, fairness must be considered at multiple levels, and classical definitions and mitigation strategies often fail to capture the complexity of real-world applications. In this thesis we treat fairness as a first-class objective in FL, on par with performance. We address four main challenges: (i) the tension between global fairness, measured across the federation as a whole, and local fairness, measured within individual clients; (ii) the extension of fairness beyond binary classification and single attributes; (iii) the simultaneous enforcement of multiple fairness constraints; and (iv) the lack of explicit control over the tradeoff between fairness and predictive performance. To this end, we develop a progression of methods that gradually expand the scope of fairnessaware FL. GLOFAIR introduces an approach that balances fairness and performance when enforcing a finite set of group fairness constraints in FL, but the trade-off is determined implicitly and cannot be fixed in advance. Building on this limitation, FairLAB is developed in a centralized setting. It extends the methodology to multiple and intersectional constraints and introduces the idea of a performance budget, which allows practitioners to decide beforehand how much accuracy they are willing to trade for fairness. In this way, FairLAB transforms the implicit compromise of GLOFAIR into a tunable design choice. We then present FeDist, a strategy for transferring knowledge across models in FL. Although focused on performance rather than fairness, it provides the technical building block that is later extended to fairness-aware training. Finally, FedFairLAB combines the principles of FairLAB and FeDist to enforce multiple, intersectional, and multiclass fairness constraints under realistic federated conditions. By doing so, it provides fairness guarantees while maintaining strong predictive performance and explicit control over performance budgets. This work advances the path towards trustworthy FL, embedding fairness into the learning process and complementing the inherent data-locality benefits of the FL paradigm to enable responsible deployment in high-stakes domains.