WP3 – Federated Learning
Federated Learning (FL) is a decentralized Machine Learning paradigm that enables multiple devices or agents to collaboratively train models without directly sharing their local data. This approach is particularly relevant in privacy-sensitive domains such as healthcare and finance, where data sharing is often constrained. Despite its rapid development, a substantial gap remains between the empirical success of FL algorithms and their theoretical understanding. By bringing a control-theoretic perspective to FL, we build novel methodologies supported based on rigorous mathematical analysis and solid statistical foundations.
Task 3.1 – Learning together without sharing the data
Federated Learning allows several participants to train a common model while keeping their data locally. Operator-splitting techniques offer a mathematical framework for coordinating this collaboration when participants hold different samples, different features or even use different models.
Task 3.2 – What does a federated model reveal about its data?
Keeping data locally does not automatically guarantee privacy: information exchanged during training may still reveal part of the original data. Observability and control-theoretic tools can help quantify this information leakage and understand when private data could potentially be reconstructed.
Task 3.3 – Federated Learning under attack
A malicious participant can deliberately inject false information into the collaborative training process. We investigate how such attacks affect the final model and how Federated Learning can be made more resilient.
Task 3.4 – Federated Learning at scale
Real federated systems may involve anything from a few organizations to thousands of mobile or IoT devices. Stochastic algorithms can make this large-scale collaboration computationally feasible while retaining reliable convergence and learning performance.
Publications
Liu K., Wang Z., & Zuazua E. (2026). Nonlinear Equilibrium Transitions in a Potential Game Model for Federated Learning. Phys. D: Nonlin. Phenom., 495, 135288. arXiv:2411.11793
Wang Z., Song Y., & Zuazua E. (2026). Approximate and Weighted Data Reconstruction Attack in Federated Learning. IEEE Trans. Big Data, 12(5), 1606-1617arXiv:2308.06822
Song, Y., Wang, Z., & Zuazua, E. (2025). FedADMM-InSa: An inexact and self-adaptive ADMM for federated learning. Neur. Netw., 181, 106-772. arXiv:2402.13989 , FAU CRIS
Morales, R. & Biccari, U. (2025). A Multi-Objective Optimization framework for Decentralized Learning with coordination constraints. Submitted. arXiv:2507.13983