Federated Logistic Regression with Mix Partitioned data#

What is Mix Partitioned Data#

Mix partitioning descibes a senario with horizontal and vertical partitioning both. An example of mix partitioned data is as follows. A and B hold same samples but different parts of features seperately, while C/D/E share same features but hold different samples.

mix.png

Partitioning#

The data will be partitioned into several vertical partitions firstly as follows.

mix_partition.png

The vertical data 0 is consist of (data0 of A, data0 of B, data of C, label0) corresponding to same sample, and so on for vertical data 1 and 2.

Algorithm#

SecretFlow provides SDG based Federated Logistic Regression to handle mix partitioned data, and each participant holds its weights.

The high level overview of the algorithm:

  1. Vertical federated logistic regression inside partitions of the same vertical data.

  2. Horizontal federated logistic regression between vertical data.

The detailed steps of the algorithm:

(suppose A holds label)

mix_lr.png