【一种数据异构的FL迁移学习框架】Privacy-preserving Heterogeneous Federated Transfer Learning

原文：Privacy-preserving Heterogeneous Federated Transfer Learning

本文讲了数据异构（区别于模型结构异构）下的 FL 训练框架，此外还有隐私保护的作用。这里的数据异构是指不同的 feature spaces，因此本文使用了迁移学习进行解决。

数据异构 heterogeneous data

数据异构有两种情况：homogeneous feature space or
heterogeneous feature space with instance co-occurrence.
因此存在两个问题：limited common features and the lack of aligned instances.
为了解决这个问题，本文提出了 heterogeneous federated transfer learning (HFTL), to enable federated learning to deal with heterogeneous feature spaces using transfer learning.

迁移学习 Transfer Learning

迁移学习使相关领域的知识得以贯通。
迁移学习，根据输入样本的同构和异构，可以分为 homogeneous transfer learning 和 heterogeneous transfer learning.

In homogeneous transfer learning (a.k.a domain adaptation), the source domain and target domain only differs in marginal distribution [2], [13], [23].
Heterogeneous transfer learning takes the variations of feature sets into consideration and is more challenging.

Model

讲一下输入特征的划分：common feature space & unique feature space


对于安全的定义是：

实现

看一下是如何实现 FL 数据异构学习的：

但是本文还涉及了很多