Research

My current research studies federated learning through the lenses of optimization, control, and dynamical systems. The central goal is to design algorithms that remain efficient, fair, and robust when data and computational resources are distributed across heterogeneous clients.

Federated Learning and Distributed Optimization

I work on optimization algorithms for federated learning, including ADMM-based methods and adaptive aggregation rules. My work on FedADMM-InSa develops an inexactness criterion and a self-adaptive penalty scheme to reduce client workload while improving robustness under heterogeneous data and systems.

Fairness and Client Contribution

I study fairness-aware aggregation in federated training, with a focus on how client contributions should be evaluated and incorporated into the global update. Recent projects include trajectory Shapley value methods and conflict-resolved aggregation for reducing client-level performance disparities.

Privacy and Reconstruction Attacks

I also investigate privacy risks in federated learning. My work on approximate and weighted data reconstruction attacks studies how private training data can be reconstructed from multi-step model updates in realistic FedAvg settings, and how architecture-aware loss weighting can improve reconstruction quality.

Dynamical Systems and Game-Theoretic Models

Another line of work models client participation and training effort through potential games and nonlinear equilibrium analysis. This connects federated optimization with dynamical systems, Nash equilibria, and incentive-aware client behavior.