A carregar...
Publicação
Attack-Resilient Federated Machine Learning at Nodes
| Nome: | Descrição: | Tamanho: | Formato: | |
|---|---|---|---|---|
| 12.06 MB | Adobe PDF |
Autores
Orientador(es)
Resumo(s)
Federated learning (FL) emerged as a solution for distributed machine learning, enabling the
training of models from decentralized data (e.g., stored in clients’ devices) without compromising
their privacy. Its distributed nature makes it vulnerable to attacks from adversaries, resulting in
possible degraded inference behavior for targeted examples or general misclassification of dataset
instances. Hence, the implementation of security mechanisms to prevent nefarious actions from
impacting the global model is critical to ensure that FL is a safe environment to train a model
collaboratively while assuring clients’ data privacy.
Evaluations of node attacks in current FL literature require complex modifications to create
or integrate new attacks, which entail considerable effort. Therefore, in this work, we propose a
tool called FADO that provides researchers and industry professionals with a platform to quickly
evaluate FL scenarios under various types of attacks. FADO is scalable, capable of executing FL
scenarios with thousands of clients while ensuring realism and a high degree of customizability.
Users can integrate their implementations of different FL components into FADO with minimal
complexity, thanks to its interfaces that facilitate the loading of external modules.
Additionally, we introduce a novel backdoor attack called BLARE, which manipulates the
attacker’s model parameters during training to make the backdoor more persistent in the global
model, even after the attacker stops injecting malicious updates. We evaluated this backdoor attack
using FADO with two datasets: CIFAR10 and CIFAR100. Several experiments were conducted in
which we varied specific hyperparameters to understand their impact on BLARE’s effectiveness.
We conclude that FADO and BLARE significantly contribute to the understanding of node
attacks in FL, offering valuable insights to improve the robustness and security of FL systems.
Descrição
Tese de Mestrado, Segurança Informática, 2024, Universidade de Lisboa, Faculdade de Ciências
Palavras-chave
Aprendizagem federada Privacidade Ataque backdoor Segurança Inteligência artificial Teses de mestrado - 2024