云鉴权>> Membership_Inference_Attack_DP>> 返回
项目作者： sharmi1206

项目描述 ：
Testing membership inference attacks on Deep learning models (LSTM, CNN);
高级语言： Python
项目地址: git://github.com/sharmi1206/Membership_Inference_Attack_DP.git


Membership_Inference_Attack_DP

Acknowledgemnet and Reference : This code relies on the research work done as https://www.biorxiv.org/content/10.1101/2020.08.03.235416v1.full. and its github https://github.com/work-hard-play-harder/DP-MIA

This repository adds an extra feature of testing privacy related ML attacks by building multi-class classificatuion models using Differential privacy.

This repository contains 2 models CNN and LSTM trained using Multi-class classification problem using DIfferential Privacy (with loss function as SparseCategoricalCrossentropy)

Requirements

Python 3.5 or higher
TensorFlow 1.14 or 1.15

pip install tensorflow-privacy

Steps to execute

  1. After downloading the code , fill the path =’’ with your current direectory path
  2. Dataset is already present with the cloned repository , does not need download. It uses daata from http://archive.ics.uci.edu/ml/datasets/Adult
  3. LSTM and CNN (conv1D) models are used to train the shadow and attack model

alt text

  1. L1 (Kernel Regularization parameter) can be tweaked to diminish the effect of overfitting and lower the attack accuracy.
  2. Both the models are trained with Differential privacy using DPGradientDescentGaussianOptimizer