This survey paper provides a comprehensive overview of the privacy landscape in federated learning (FL). It begins by introducing the federated learning process and the key considerations for ensuring operational integrity and security. The paper then delves into the different variants of FL, including cross-device, cross-silo, federated transfer learning, and hierarchical FL, as well as the implications of IID and non-IID data distributions on privacy and learning.
The core of the survey focuses on the threat model for privacy attacks in FL, categorizing them into central server threats (honest-but-curious and malicious servers) and client-side threats (honest-but-curious and malicious clients). The paper then dives deep into the different types of privacy attacks:
Data reconstruction attacks: These attacks aim to directly reconstruct the client's private data or a representation of it. The survey covers optimization-based, linear layer leakage, GAN-based, and other data reconstruction attacks, highlighting their requirements, success, and limitations.
Membership inference attacks: These attacks infer whether a particular sample was used in the training of the target model.
Property inference attacks: These attacks learn about sensitive properties within the training set, such as race, gender, or age.
Model extraction attacks: These attacks aim to steal the functionality of the target model, including its parameters and hyperparameters.
The paper also discusses the various defense mechanisms proposed to mitigate these privacy attacks, including differential privacy, secure aggregation, homomorphic encryption, and trusted execution environments. It highlights the limitations and drawbacks of each defense approach.
Moving beyond the technical landscape, the survey examines the real-world applications of FL across different industries, such as healthcare, finance, and IoT/edge computing. It showcases how FL is being leveraged to address privacy concerns while enabling collaborative learning.
Finally, the paper delves into the evolving policy landscape surrounding data privacy, discussing regulations like GDPR, HIPAA, and emerging legislation in the US and EU. It emphasizes the need for robust privacy-preserving techniques, like FL, to align with these regulatory requirements and enable the widespread adoption of FL in sensitive domains.
To Another Language
from source content
arxiv.org
Ключові висновки, отримані з
by Joshua C. Zh... о arxiv.org 05-07-2024
https://arxiv.org/pdf/2405.03636.pdfГлибші Запити