README.md

Differential Privacy for Gaussian Processes

paper maybe : https://arxiv.org/pdf/1606.00720.pdf

or: https://pmpml.github.io/PMPML16/papers/PMPML16_paper_4.pdf 。（by WT）

Installation

Clone the repo and use pip to install;

git clone https://github.com/lionfish0/dp4gp.git
cd dp4gp
pip install -e .

Usage

Fairly simple to use. Simply build a normal Gaussian Process regression model in GPy, using any kernel you would like. Then pass it to the dp4gp method of your choice.

from dp4gp import dp4gp
import numpy as np
import GPy
import matplotlib.pyplot as plt
%matplotlib inline

X = np.arange(0,10,0.1)[:,None]
Y = np.sin(X)+np.random.randn(len(X),1)*0.4
kern = GPy.kern.RBF(1.0,lengthscale=2.0,variance=1.0)
model = GPy.models.GPRegression(X,Y,kern,normalizer=None)
#model = GPy.models.SparseGPRegression(X,Y,kern,normalizer=None)
model.Gaussian_noise = 0.3

dpgp = dp4gp.DPGP_cloaking(model,ac_sens,epsilon=1.0,delta=0.01)
#dpgp = dp4gp.DPGP_inducing_cloaking(model,ac_sens,epsilon=1.0,delta=0.01)

dpgp.plot(plot_data = True,extent_lower={0:0},extent_upper={0:20},Nits=100,confidencescale=[1.0]);
plt.ylim([-5,5])