Concepts de base
Federated Computing enables collaborative data processing and machine learning across distributed devices without compromising individual data privacy.
Résumé
The content provides a comprehensive survey on Federated Computing (FC), which is a decentralized approach to data processing and machine learning. FC addresses the challenges of traditional centralized computing models, such as data security breaches and regulatory hurdles, by enabling collaborative processing without compromising individual data privacy.
The key aspects covered in the survey are:
Definition and Distinction of Federated Learning (FL) and Federated Analytics (FA):
FL focuses on collaborative machine learning model training, while FA leverages statistical operations on distributed data.
Both fall under the broader umbrella of FC, which aims to extract insights from distributed data sources without disclosing raw data.
FC System Components and Extensions:
Basic building blocks: Client selection, aggregation strategy, and communication protocol.
Extensions: Privacy-enhancing techniques (e.g., differential privacy, secure multi-party computation) and compression methods to optimize network utilization.
FC System Architectures:
Centralized, hierarchical, and peer-to-peer architectures with varying levels of decentralization.
FC Scenarios and Challenges:
Model-centric vs. data-centric, horizontal vs. vertical, and cross-device vs. cross-silo scenarios.
Key challenges include improving insights/ML performance, enhancing privacy/security, and optimizing hardware/network utilization.
The survey provides a comprehensive taxonomy and framework to describe FC systems, highlighting the interplay between the basic building blocks and optional extensions. It also identifies research gaps and prevalent system configurations in the current literature.