Keskeiset käsitteet
Proposing a FedDRL method based on reinforcement learning for trustworthy model fusion in federated learning, addressing challenges of malicious models and low-quality data.
Tiivistelmä
The content introduces the FedDRL method for trustworthy model fusion in federated learning. It discusses challenges faced by conventional methods, proposes a two-stage approach using reinforcement learning, and validates the method through experiments with different scenarios and datasets.
Abstract:
Introduces challenges in conventional federated learning approaches.
Proposes FedDRL method based on reinforcement learning for trustworthy model fusion.
Introduction:
Discusses the importance of deep learning technologies in various industries.
Highlights the need for collaborative data analysis while preserving user privacy.
Related Work:
Reviews existing research on federated learning algorithms like FedAvg, Scaffold, FedProx, etc.
Discusses challenges related to Non-IID data distribution in federated learning models' convergence speed.
Method:
Defines the problem statement and proposes a two-stage approach using reinforcement learning for client selection and weight assignment.
System Design:
Outlines the framework workflow for trustworthy federated learning using staged reinforcement learning.
Experiment:
Conducts experiments on malicious client attacks, low-quality model fusion, and hybrid scenarios across different datasets to evaluate the FedDRL method's performance.
Tilastot
Reinforcement learning (RL) employs a trial-and-error strategy.
Continuous training is required for sample collection through environmental interaction.
Traditional single-agent reinforcement learning training approaches can be time-consuming.