Locally Differentially Private Distributed Deep Learning via Knowledge Distillation (LDP-DL)

• A preprint version of our paper: Link here

• This is a sample code of our LDP-DL paper, including all the necessary implementations: 1) DataPartation.py: split the data for data owners and data user; 2) TeacherModel.py: train the teacher models using data owners' private data; 3) StudentModel_AQS.py: train the student model via active query sampling and knowledge distillation using data user's public data.

• The splited data and trained teacher models are also provided. Due to the limitation of capacity, please download from this link:

Requirements

• Pytorch 1.4.0, numpy 1.18.1, scipy 1.4.1, cuda 10.0, joblib 0.14.1

Quick Start (only tested using PyCharm on Ubuntu, Python 3.7)

• Download or clone the whole repository.
• Import our code as a project into PyCharm.
• Run DataPartation.py to split the data.
• Run TeacherModel.py to train teacher models.
• Run StudentModle_AQS.py to train and evaluate student model.

Contact info: [email protected], [email protected]