A carregar...
Publicação
Understanding Synchronous vs. Asynchronous Federated Learning in FADO
| Nome: | Descrição: | Tamanho: | Formato: | |
|---|---|---|---|---|
| 3.05 MB | Adobe PDF |
Autores
Orientador(es)
Resumo(s)
Federated Learning (FL), a decentralized machine learning paradigm, has received increasing attention in recent years. While early FL methodologies focused on synchronous communication, recent research has shown that hybrid and asynchronous algorithms can mitigate key limitations such as straggler effects and synchronization delays. This thesis explores asynchronous and hybrid FL algorithms by evaluating their performance in realistic scenarios and extending the Federated Learning Attack and Defense Orchestrator (FADO) to support these communication methodologies. The FADO framework, originally designed for synchronous FL, was restructured to emulate client-server interactions under asynchronous conditions. This modification expanded the framework capabilities to support diverse scenarios, aggregation algorithms, attack vectors, and defense mechanisms. Key implementation efforts included a new client-server emulation, capable of multi-system deployment, a modular orchestration workflow for selecting aggregation, participation, and communication strategies, and the integration of real-time network latency based on measured round-trip delays between server and client locations. Evaluation focused on comparing algorithm performance baselines and exploring how various FL algorithms behave under conditions of data and device heterogeneity. For data heterogeneity, experiments varied the class distribution of training data among clients with the CIFAR-10 dataset to simulate non-IID conditions. For device heterogeneity, measured network latency and straggler behavior was introduced across clients. Results showed that asynchronous algorithms achieved faster convergence and more efficient resource utilization in device heterogeneity scenarios, though they introduced challenges such as delayed updates. These results demonstrate the practical relevance of the framework and the capabilities of its modularity structure when creating different, realistic and variable conditions. This thesis contributes a restructured version of FADO, a modular, extensible framework for evaluating any type of FL algorithms, along with insights into their performance metrics and tradeoffs. This framework can serve as a basis for future research into new types of FL and their defenses.
Descrição
Tese de Mestrado, Segurança Informática, 2025, Universidade de Lisboa, Faculdade de Ciências
Palavras-chave
Federated Learning Asynchronous FL Privacy FL Framework FADO