Hetero FTL¶

Now, hetero federated transfer learning is refactorized based on fate-1.5. And additional test datasets are offered.

This folder contains code for implementing algorithm presented in $Secure Federated Transfer Learning$ ..

Our FTL algorithm is trying to solve problem where two participants -host and guest - have only partial overlaps in the sample space and may or may not have overlaps in the feature space. This is illustrated in Figure 1. Our objective is to predict labels for host as accurately as possible.

Figure 1： Federated Transfer Learning in the sample and feature
space

for two-party problem

Our solution employs an architecture of two layers: local layer and federation layer.

Figure 2: Architecture of Federated Transfer
Learning

In the Local layer, both guest and host exploit a local model for extracting features from input data and outputting extracted features in the form of numerical vectors. The local model can be CNN for processing images, RNN for processing text, dense layer for processing general numerical vectors and many others. Currently users can define their own local layers(including nn structure, optimizer, learning rate) in the algorithm conf.

The federation layer is for the two sides exchanging intermediate computing components and collaboratively train the federated model by minimizing target loss and alignment loss.

In current version, we get rid of the arbiter and still guarantee data/model privacy. For detail on how this can be achieved, please refer to $Secure Federated Transfer Learning$ ..

We support two mode for the FTL algorithms: plain mode and encrypted mode. In plain mode, data are computed and transferred in plaintext, while in encrypted mode intermediate results will be encrypted using Paillier.

What is more, in the latest Hetero FTL, we add an option: communication efficient mode. Once communication efficient mode is enabled, for every epoch, intermediate components are preserved and are used to conduct several local model weights updates, thus communication cost is reduced.

Features¶

Support plain/encrypted mode
Support local layer define / optimizer define
Support communication-efficient mode.

Applications¶

Now Hetero FTL only supports binary classification.

Quick Start¶

Now you can start hetero-ftl like other algorithms in FATE. Please refer to the "examples" folder.

Last update: 2021-11-15