Python Toolbox for Robustness Verification of Quantum Classifiers

This toolbox is implemented on Python and can verify the robustness of quantum machine learning algorithms with respect to a small disturbance of noises, derived from the surrounding environment.

This toolbox makes use of Numpy and an SDP solver — CVXPY: Python Software for Disciplined Convex Programming.

1. The installation of VeriQ requires BLAS and LAPACK. Cmake and pip3 are also needed.

sudo apt install -y libblas-dev liblapack-dev cmake python3-pip

2. Because the default version of Python in Ubuntu 18.04 is Python3.6, Numpy should be installed first.

pip3 install numpy

3. Install CVXPY.

pip3 install cvxpy

4. Besides, our demonstration file batch_check.py uses a Python library PrettyTable for printing a format table summarizing the numerical results, and matplotlib is also needed to be installed for generating visualized adversary examples for the classification of MNIST.

pip3 install prettytable matplotlib

5. Clone or download the VeriQ toolbox from VeriQ. All files must be saved in the same location.

git clone https://github.com/j68249959/VeriQ

To implement robustness verification on VeriQ, we assume that the user has already trained a quantum classifier which consists of a quantum circuit with a measurement at the end. The quantum circuit and the training data have been saved as a NumPy data file.

The user can use the following script to run VeriQ for robustness verification of quantum classifiers.

python3 batch_check.py <data_file> <robustness_unit> <experiment_number> <state_flag>

There are four arguments are inputted by users. The first one <data_file> is a NumPy data file that consists of a (well-trained) quantum classifier and corresponding training dataset. The NumPy data file can be directly obtained by the data of the classifiers trained on the platform --- Tensorflow Quantum of Google. The second argument <robustness_unit> is the unit of the robustness verification parameter. The third argument <experiment_number> represents the number of robustness verification with increasing robustness verification parameters by unit <robustness_unit>. For example, if <robustness_unit> and <experiment_number> are 1e-3 and 4, respectively, then 1e-3, 2e-3,3e-3, 4e-3-robustness of the quantum classifier to be checked one by one. The last one <state_flag> indicates the considering quantum data in robustness verification, where the value of <state_flag> is mixed or pure referring to mixed states and pure states, respectively.

For simplicity, the user can try the following commands in bash for robustness verification of four quantum classifiers.

1. Quantum Bits Classifications

python3 batch_check.py binary_cav.npz 1e-3 4 mixed

2. Quantum Phase Recognition

python3 batch_check.py phase_recog_cav.npz 1e-4 4 mixed

3. Cluster Excitation Detection

python3 batch_check.py excitation_cav.npz 1e-4 4 mixed

4. The Classification of MNIST

python3 batch_check.py mnist_cav.npz 1e-4 4 pure

The user can use the following script to obtain the adversarial examples of MNIST classification, which are found by VeriQ. The results are generated into .png files.

python3 generate_adversary.py

The models of the above classifiers can be found in the Figures file.

1. Quantum Bits Classifications
2. Quantum Phase Recognition
3. Cluster Excitation Detection
4. The Classification of MNIST
5. Adversarial Examples