This year our team member Ziqi Wang will present his PhD Thesis on:
“Federated Learning under Heterogeneity: Efficiency, Fairness, and Safety”
Advisor:
• Prof. Enrique Zuazua, FAU – Friedrich Alexander-Universität Erlangen-Nürnberg (Germany)
Abstract. Machine learning has become central to modern science and technology, and its performance depends heavily on diverse training datasets. However, such data are often distributed across devices and institutions, making centralized collection impractical due to privacy constraints. Federated learning (FL) addresses this issue through a client–server system in which clients retain raw data locally while a server coordinates training of a shared model. Mathematically, empirical risk minimization in this setting requires coordinating clients with heterogeneous data and computational resources, giving rise to intertwined challenges of efficiency, fairness, and safety. This thesis studies these challenges through communicated client updates: how they are produced and supplied, how they are valued and aggregated, and what private information they reveal.
The first part studies client-side efficiency and participation, where heterogeneous computational capabilities and costs shape the production and supply of client updates. We propose an inexact ADMM-based FL method that adapts local workload through a first-order residual criterion instead of prescribed local efforts. Under smoothness and strong convexity, primal–dual convergence and vanishing local optimality residuals are established. Participation fairness is then modeled as a strategic effort game under heterogeneous costs. The resulting weighted potential game admits a Nash equilibrium. Under stationary quadratic costs, equilibrium uniqueness, a critical transition in client effort, and cyclic best-response convergence are further characterized.
The second part investigates server-side valuation and aggregation, in which fairness concerns both clients’ influence and the resulting client-level performance. We develop FedTSV, a contribution-aware aggregation method that measures marginal contributions through trajectory Shapley values relative to a validation-derived reference update and converts them into adaptive aggregation weights. Its Monte Carlo estimator is unbiased and satisfies a concentration bound. Since scalar weighting does not control aggregation geometry, we further propose CRAFT, a method that resolves update conflicts through a reference-anchored constrained least-squares problem with a closed-form pseudoinverse solution. We also provide common-descent and finite-time stationarity guarantees.
The third part turns to safety and explores what private information may be inferred from the communicated client updates. In an honest-but-curious server setting, we propose an adaptive weighted reconstruction attack that targets the final locally trained model rather than requiring access to intermediate local states. It approximates unobserved local training trajectories by interpolation and reconstructs client data through layer-wise weighted update matching, with weights selected by Bayesian optimization. Experiments show that substantial information can remain recoverable even when raw data stay local and only the final locally trained model is communicated.
Taken together, this thesis links client-side computation and participation, server-side valuation and aggregation, and privacy exposure through communicated updates within a coherent treatment of efficiency, fairness, and safety in FL under heterogeneity.
WHEN
TBA, 2026 HH:HH (Berlin time)
WHERE
On-site. FAU
Board of examiners
|| You might like: Upcoming events