Machine learning continues to shape consumer technologies and industrial applications in many ways. Training these systems usually requires collecting data in one place, which is often challenging or simply impossible. Federated learning (FL) is a newer approach that avoids this: instead of sending data to a central server, each device trains its part of the model locally and shares only small updates based on the local data. The server then combines these updates into an improved model, which is later disseminated across the participating devices.
This setup allows many devices to learn together without exposing their data, but it also introduces new challenges. The central challenge comes from the fact that real-world systems rarely operate under uniform conditions. Devices may differ in computing power, data varies from one user to another, and network delays are often unpredictable. Conventional federated learning methods, which are designed for homogeneous, well-synchronized environments, often fail under these conditions. This can cause the system to slow down, drop participants, or produce uneven model quality across users.
New lightweight FL algorithms for heterogeneous devices
Aleksei Ponomarenko-Timofeev’s doctoral dissertation at Tampere University addresses this core problem by developing novel lightweight FL algorithms for heterogeneous devices. A central idea in the thesis is personalization, i.e., the ability for each device to learn a version of the model that fits its own data, while still contributing to the global model shared across all participants and collective training.
This approach improves both fairness and performance: devices with highly diverse data benefit from learning the common patterns in the data, while their unique data characteristics remain represented in the final models. Experiments show that the proposed methods reduce the training delays by up to 20% compared to state-of-the-art FL methods, while achieving more consistent accuracy.
The dissertation also introduces new ways to handle “stragglers”, i.e., devices that fall behind because of slow hardware or unstable networks. Instead of discarding their contributions, the proposed policies incorporate partial or delayed updates. This leads to faster convergence and more stable model quality across users, while maintaining fairness.
Privacy is preserved through differential privacy mechanisms that incorporate carefully controlled noise into the updates of individual users. The work shows how to apply these tools without significantly weakening the quality of the final models.
The dissertation presents a framework for personalized, resource-aware, and privacy-preserving FL. The results are directly relevant for scenarios where data cannot be centralized, and devices operate under strict hardware and privacy constraints. The potential applications may be telecommunications, healthcare, and large-scale Internet-of-Things (IoT) deployments.
Public defence on Friday, 12 December 2025
The doctoral dissertation of M.Sc. Aleksei Ponomarenko-Timofeev, in the field of Communications Engineering, titled Personalized Support Vector Machines for Privacy-Preserving Federated Learning will be publicly examined at the Faculty of Information Technology and Communication Sciences at Tampere University at 12:00 on Friday, 12.12.2025, at Hervanta campus, Tietotalo, TB109 (Korkeakoulunkatu 1, Tampere).
The Opponent will be Dr. Nicola Marchetti of Trinity College of Dublin, Ireland. The Custos will be Dr. Sergey Andreev of Tampere University, Finland.
The doctoral dissertation is available online.
The public defence can be followed via remote connection.