This website requires JavaScript.

Federated Learning with Imbalanced and Agglomerated Data Distribution for Medical Image Classification

Nannan WuLi YuXin YangKwang-Ting ChengZengqiang Yan
Jun 2022
摘要
Federated learning (FL), training deep models from decentralized data withoutprivacy leakage, has drawn great attention recently. Two common issues in FL,namely data heterogeneity from the local perspective and class imbalance fromthe global perspective have limited FL's performance. These two couplingproblems are under-explored, and existing few studies may not be sufficientlyrealistic to model data distributions in practical sceneries (e.g. medicalsceneries). One common observation is that the overall class distributionacross clients is imbalanced (e.g. common vs. rare diseases) and data tend tobe agglomerated to those more advanced clients (i.e., the data agglomerationeffect), which cannot be modeled by existing settings. Inspired by real medicalimaging datasets, we identify and formulate a new and more realistic datadistribution denoted as L2 distribution where global class distribution ishighly imbalanced and data distributions across clients are imbalanced butforming a certain degree of data agglomeration. To pursue effective FL underthis distribution, we propose a novel privacy-preserving framework named FedIICthat calibrates deep models to alleviate bias caused by imbalanced training. Tocalibrate the feature extractor part, intra-client contrastive learning with amodified similarity measure and inter-client contrastive learning guided byshared global prototypes are introduced to produce a uniform embeddingdistribution of all classes across clients. To calibrate the classificationheads, a softmax cross entropy loss with difficulty-aware logit adjustment isconstructed to ensure balanced decision boundaries of all classes. Experimentalresults on publicly-available datasets demonstrate the superior performance ofFedIIC in dealing with both the proposed realistic modeling and the existingmodeling of the two coupling problems.
展开全部
图表提取

暂无人提供速读十问回答

论文十问由沈向洋博士提出,鼓励大家带着这十个问题去阅读论文,用有用的信息构建认知模型。写出自己的十问回答,还有机会在当前页面展示哦。

Q1论文试图解决什么问题?
Q2这是否是一个新的问题?
Q3这篇文章要验证一个什么科学假设?