Adversarial attack against label propagation

Source code of the paper Influence-Driven Data Poisoning in Graph-Based Semi-Supervised Classifiers accepted at CAIN 2022.
Author of the code: Adriano Franci.

Setup

Tested under python 3.6.8 and 3.7.2.
Please extract the svm_light MNIST dataset version archives (can be found in data folder)

Environment install:

• python3 -m venv venv/
• source venv/bin/activate
• pip install -r requirements.txt

Run

Define experiement parameters on the config.yaml file. (applicable to research question 2,3,4 as RQ1 has static config)

Run the experiment using the corresponding main_rq*.py script. The results outputs are as follow :

• main_rq1.py : single file per dataset : transductive_accuracy_exhaustive_lp_ls_${dataset}_${p_labelled} representing the accuracy decrease (relative decrease) by flipping each of the labeled samples.

• main_rq2.py : files : {inductive_algo}_{ssl_name}_{dataset}_{p_labelled} representing the absolute accuracy for inducive algo = [rfc,mlp] ssl = [lp,ls] dataset = [mnist,cifar,rcv1] for three different unlabelled proportions p_labelled = [5%,15%,25%] given 5 flip budget = [0%,5%,10%,15%,20%]

• main_rq3 : same as rq but with attack algorithm = [random,greedy,probabilistic] with time efficiency report.

Plots helpers functions can be found in representation/main.py

Non exhaustive Requirements

cycler==0.10.0
decorator==4.4.2
imageio==2.9.0
joblib==0.16.0
kiwisolver==1.2.0
matplotlib==3.3.0
networkx==2.4
numpy==1.19.1
Pillow==7.2.0
pyparsing==2.4.7
python-dateutil==2.8.1
PyWavelets==1.1.1
scikit-image==0.17.2
scikit-learn==0.23.2
scipy==1.5.2
six==1.15.0
threadpoolctl==2.1.0
tifffile==2020.7.24