Muffliato

This private repository contains the code for the paper Muffliato: Peer-to-Peer Privacy Amplification for Decentralized Optimization and Averaging.

Requirements

The following packages are used: numpy matplotlib scikit-learn scipy networkx tqdm typer pathlib

Data

The Facebook ego graph can be downloaded here: https://snap.stanford.edu/data/ego-Facebook.html and should be placed in a folder named 'facebook' in the project folder. The Housing dataset should be downloaded via the 'data' script, but can be manually retrieved there in case of problem: https://www.openml.org/d/823

To reproduce the figures of the main paper

Fig a

Run figa.py

Fig b

After downloading the Facebook Ego dataset as indicated before, run figb.py

Fig c

Run gd.py to generate the data (takes several minutes, you can reduce the number of trials by adding '--n_trials 2 < INT< 10'). The files are saved in a folder 'result' Run draw.py to generate the figure b on two seperate runs simulated with the gd.py files. Beware that you should specifiy which runs you want to use.

In command lines, to reproduce with the same parameters as in figure b

mkdir result
python --n_trials 10 --n_nodes 2000 --prefix "big" gd.py
python --n_trials 10 --n_nodes 4000 --prefix "huge" gd.py
python draw.py

To reproduce the figures of the supplementary material

Run fig2.py --graph-name hypercube/ER/geometric/grid to reproduce the figure 2. If no parameter is given, hypercube is used. You can also modify the number of nodes of those graphs.

Run fig3.pyto reproduce the Figure 3.

Run fig4.py to reproduce the Figure 4. Although no direct command is provided, you can easily modify the number of nodes, the positions of the node and the radius and starting node via variables n, pos, and u (line 24 to 30).

Run fig5.py to reproduce the Figure 5, after downloading the Facebook Ego graph as indicated before.

To generate a run of figure 6, run versatile.py by putting the seed in parameter and a prefix. This generates the two npy with the privacy loss and the mean square error at each step. By putting the list of the prefix generated, the plot can be done using fig6.py.

Short description of the files

Core of the project

graphutils.py Several useful functions such as building the gossip matrix and computing the spectral gap
muffliato.py Muffliato algorithm, non accelerated and accelerated version
ERsampled.py privacy leakage for Erdös Renyi graph drawn randomly (theorem 4)

For synthetic experiments

figa.py for comparing several graphs as described in the main text
fig2.py for visualizing the privacy loss of a fixed node on several graphs
fig3.py for the privacy loss of the exponential graph
fig4.py for the privacy loss as function of euclidean distance for random geometric graphs.
fig6.py for visualizing the privacy with dropout
versatile.py for generating runs with dropout

For Fb experiments

figb.py generates the current Fb drawing, you can change the ego variable to run simulation on other graphs
fig5.py to generate the full gallery of Ego graphs.

For the Gradient Descent part

data.py Used for preprocessing the Housing dataset for the logistic regression
gd.py Main file for the logistic regression
private.py logistic regression utils and construction of Muffliato Logistic Regression algorithm

totilas/muffliato