Defending Adversarial Attacks in MIMO Systems

This repository hosts code used to obtain results in our paper: Defending Adversarial Attacks on Deep Learning Based Power Allocation in Massive MIMO Using Denoising Autoencoders

Repository Structure

• data/: This folder contains the testing data set. Our work uses the publicly available Power Allocation in Multi-Cell Massive MIMO dataset. To download the training data set for our experiments, download the multi_cell.zip file, unzip and copy the file named dataset_maxprod.mat into the data/ folder in this repository.
• saved_nn_models/: This folder contains saved neural network models from our experiments. Load these models to obtain the same results we showed in our paper. model_1/ and model_2/ sub-directories contain the saved models for model architecture 1 and model architecture 2 respectively, which are detailed in our paper.
• src/ folder contains:
  • attacks.py that implements the adversarial attacks we use.
  • model architecture and training scripts: baseline.py, dae_training.py and adv_regressor.py.
  • result evaluation scripts: eval_networks.py to evaluate semi-whitebox experiments and eval_blackbox.py to evaluate blackbox experiments.
  • Makefile that facilitates running the experiments.
• requirement.txt: A snapshot of the Python package versions the experiments were run with.

Get Started

How to Use Saved Models to Reproduce Our Results

1. Ensure that you're in the src/ folder:

$ pwd
DAE_for_adv_attacks_in_MIMO/src

2. Use the Makefile to run semi-whitebox experiments:

make eval_all

3. Use the Makefile to run blackbox experiments:

make eval_blackbox_all

How to Re-train the Models

1. Ensure that you're in the src/ folder:

$ pwd
DAE_for_adv_attacks_in_MIMO/src

2. Ensure that you have downloaded the dataset zip file from the dataset website, and have copied the training set into the data/ folder as /data/dataset_maxprod.mat.
3. To re-train the baseline DL model:

make baseline_all

4. To re-train the DAE defense model:

make dae_all

5. To re-train the adversarially-trained regressor model:

make adv_regressor_all

Package Requirements

To ensure success running of the program, the versions Python packages we used are listed in requirements.txt. To align the versions of your packages to this file, simply run:

pip install -r requirements.txt