المفاهيم الأساسية
ASYN2F is an effective asynchronous federated learning framework with bidirectional model aggregation, achieving higher performance compared to existing techniques.
الملخص
The paper introduces ASYN2F, an asynchronous federated learning framework with bidirectional model aggregation. It addresses the issue of training delay caused by heterogeneous training workers, leading to obsolete information. ASYN2F allows for asynchronous model aggregation at the server and local model aggregation at the workers, improving performance. The framework considers practical implementation requirements like cloud storage and message queuing protocols. Extensive experiments show superior performance and scalability of ASYN2F.
Introduction
Federated learning addresses challenges of distributed data learning.
Framework design focuses on bidirectional model aggregation.
Related Work
Various asynchronous algorithms proposed for federated learning.
Methods for global model aggregation at the server.
Framework Design and Model Aggregation Algorithms
ASYN2F architecture includes server, worker, storage, and queue components.
Algorithm 1 outlines the training process, with global model aggregation at the server.
Algorithm 2 details global model aggregation based on data quality and loss value.
Algorithm 3 describes local model aggregation at workers incorporating global model updates.
Privacy Preservation Analysis
ASYN2F framework ensures data privacy protection for workers.
Threat model analysis for data privacy in training scenarios.
Experiments
Performance comparison with existing techniques on CIFAR10 dataset.
Results show ASYN2F outperforms FedAvg and M-Step KAFL.
Faster convergence and lower communication cost with ASYN2F.
الإحصائيات
"ASYN2F achieves higher performance compared to existing techniques."
"Extensive experiments show ASYN2F's effectiveness and scalability."
اقتباسات
"ASYN2F achieves higher performance compared to the state-of-the-art techniques."
"The experiments demonstrate the effectiveness, practicality, and scalability of ASYN2F."