Name: EDUARDO MONTAGNER DE MORAES SARMENTO
Publication date: 30/10/2024
Examining board:
Name |
Role |
|---|---|
| LUIZ FERNANDO BITTENCOURT | Examinador Externo |
| MAGNOS MARTINELLO | Examinador Interno |
| RODOLFO DA SILVA VILLACA | Presidente |
| VINICIUS FERNANDES SOARES MOTA | Coorientador |
Summary: Federated learning presents a promising approach for distributed machine learning, par-
ticularly in privacy-sensitive and resource-constrained environments. However, challenges
related to client heterogeneity, computational limitations, and data privacy remain signifi-
cant barriers to widespread adoption. This dissertation investigates these challenges by
developing an emulation framework, MininetFed, based on asynchronous message exchange
communication to model diverse network conditions in federated learning environments,
easing the process of development and experimentation of federated learning solutions.
Additionally, we explored privacy-enhancing techniques by developing the CKKSFED
algorithm that uses Fully Homomorphic Encryption to securely cluster clients in the
federated learning process. Another privacy enhancing technique we also explored was
Differential Privacy with the FedSketch algorithm implemented with probabilistic data
structures to improve the security of the transfer of model updates between clients and
aggregation server while also improving the efficiency of communication by the merit of
compressing said updates into sketches.
Experimental results demonstrate that the proposed methods enhance the security and
improve computational efficiency of federated learning, including in scenarios involving
low-power devices and time-series data, while maintaining accuracy similar to the ones
reached by classical methods found in the literature. The findings highlight opportunities
for improving both the privacy guarantees and scalability of Federated learning systems,
offering a foundation for future developments in secure and efficient decentralized learning.
