data-poisoning-linear-systems

Code for the paper Analysis and Detectability of Offline Data Poisoning Attacks on Linear Systems.

Author: Alessio Russo (alessior@kth.se)

License: MIT

Requirements

To run the examples you need the following libraries: numpy scipy cvxpy dccp momentchi2 torch matplotlib. See also the requirements file requirements.txt.

Input poisoning attack

To run the example, run the file main.py in the folder example_input_attack. Feel free to change the standard deviation of the process noise std_w, or of the input signal std_u.

Residuals maximization attack

To run the example, run the file main.py in the folder example_residuals_maximization. Use plot_data.py to plot the results.

To use the data of the paper, first extract the data from the zipped files example_residuals_maximization\data\compressed_data.*.zip in the data folder.

Stealthy attack

All the data used in the papers can be found in the folder example_stealthy_attack\data.

To plot the data use the analyse_results.py file.
To run a new set of simulations use extract_data.py to collect the datasets, and algorithm.py to compute the poisoning attack using SLSQP.