.png){kind=icon}
.png){kind=icon}
Return
Fairness-aware loss history based federated learning heuristic algorithm
.png){kind=icon}
PRE
AI
DOI：10.1016/j.knosys.2024.111467.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Federated learning (FL) is a distributed learning paradigm for massively distributed clients. It is aimed at collaboratively training a model without data sharing which guarantees privacy of performance data. Due to the heterogeneity of clients' data in federated learning, the aggregated model might be biased which can lead to unfairness. Hence, clients' accuracy variance should be reduced which is referred to as fairness in federated learning. In the related works, optimizing fairness without sacrificing accuracy is regarded as a significant challenge. In this paper, we presented a new and optimized stochastic gradient descent based optimizer which applies loss history for updating model parameters and minimizing cost function. As a heuristic algorithm, we proposed Loss History Federated Learning (LHFed). It introduces an optimization scheme which uses a novel optimizing method by capitalizing on loss history. Rather than using the last loss of each client, the method applies clients' loss histories so as to enhance fairness. We conducted experiments to compare the proposed algorithm with FedAvg, q-FedAvg, FedFa and AFL algorithms. The experimental results of the federated datasets indicate that the LHFed algorithm outperforms the other algorithms with respect to the criteria of fairness, clients' accuracy, F1 score and convergence speed.
Keywords:
Federated learning
Fairness
Optimization scheme
Loss history
Journal
K
IF:
7.6
Papers:
1.2W
Citations:
4.5W