24April2026
10:00 Master's Defense room 85 of IC2
Topic on
Efficient Federated Learning on Edge Devices: Resource-Aware Methods under System Heterogeneity
Student
Bruno Santos Martins
Advisor / Teacher
Leandro Aparecido Villas
Brief summary
Federated Learning (FL) allows devices to collaborate in training a shared model while keeping the data locally. In the typical synchronous protocol, training occurs in rounds: a subset of clients downloads the global model, performs local optimization, and sends updates to a central server for aggregation. Although widely used, this dynamic is communication and computationally intensive, increasing bandwidth and energy consumption and potentially excluding clients with limited resources. Furthermore, heterogeneity in hardware, battery, and connectivity creates bottlenecks that determine the duration of rounds, generating idle time in faster clients and increasing the total training time. The literature has explored model heterogeneity, in which clients train submodels tailored to their capabilities under the same global model. However, many approaches maintain the submodel size constant throughout the training process. In practice, client capabilities are not always known a priori (and the set of participants may be partially unknown or vary), and effective performance can change due to network fluctuations and concurrent background loads; thus, static allocations tend to become suboptimal. In this dissertation, we propose two adaptive submodel allocation methods to reduce the impact of stragglers and high resource usage in FL under system heterogeneity. FedPulse combines communication and computation times to identify stragglers as outliers and adjusts the submodel size per client: smaller for slow clients, reducing wall time, and larger for fast clients, preserving accuracy. To balance energy, accuracy, and training time, we propose FedLoad, which uses non-sensitive contextual information to select submodel sizes based on energy efficiency, hardware and network capabilities, and learning progress. In experiments with simulated heterogeneous clients and standard image classification tasks, FedPulse reduces training time by up to 58% with less than 1% accuracy loss compared to FL with a homogeneous model. FedLoad, on the other hand, reduces the average run duration, energy consumption, and communication cost by up to 21,6%, 33,6%, and 29,5%, respectively, when compared to literature baselines.
Examination Board
Headlines:
| Leandro Aparecido Villas | IC / UNICAMP |
| Maycon Leone Peixoto | DCC/UFBA |
| Luis Henrique Maciel Kosmalski Costa | POLI/UFRJ |
| Juliana Freitag Borin | IC / UNICAMP |
Substitutes:
| Luiz Fernando Bittencourt | IC / UNICAMP |
| Bruno Yuji Lino Kimura | ICT / UNIFESP |