Research On Federated Learning Methods With Noisy Labels And Imbalanced Data

Federated learning is an emerging learning paradigm aimed at jointly learning knowledge from large-scale user nodes in a way that does not expose user privacy.However,due to the significant manpower required to obtain high-quality labeled data,user nodes inevitably carry noisy labeled data.At the same time,federated learning faces long tailed data distribution in application scenarios,where the number of data of different classes is highly imbalanced between local and global data.Both of these issues can lead to a serious decrease in model training effectiveness in federated learning scenarios.This thesis considers solving two problems simultaneously,and its main research content can be summarized as follows:Firstly,a long tailed robust and noisy resistant federated learning framework called Double Net Double Teacher(DNDT)framework is proposed.This framework is divided into two stages to process noisy labels and long tailed distribution data.In stage one,double networks are used to filter small loss samples for denoising,and in stage two,double teacher distillation is used to process long tailed distribution data.Secondly,three federated learning algorithms based on the DNDT framework,Fed Match,Fed Jo Co R,and Fed Coach,are proposed.These three algorithms adopt three different double network structures in the first stage and are implemented in conjunction with the same second stage design.Then,to address the communication cost issue of double network structures in federated learning,a low communication cost algorithm Fed Single Aug based on the DNDT framework is proposed.This algorithm uses a collaborative network simulated by a single network to filter noisy label samples,combined with a self-balanced method,effectively solving the problem of model accuracy degradation in federated scenarios with long tailed and noisy labels.Finally,two datasets with long tailed distribution and noisy labels were constructed based on CIFAR10 and Fashion MNIST.Based on these two datasets that are more in line with real-world scenarios,the performance of the four algorithms proposed in the paper based on the DNDT framework will be compared with existing federated learning algorithms.The results indicate that the Fed Single Aug algorithm has the best comprehensive performance.Subsequently,ablation experiments were conducted on the best performing Fed Single Aug to validate the effectiveness of the DNDT framework.