QuGAP: Generating Universal Adversarial Perturbations for Quantum Classifiers

Repository containing all the source code used for the paper Generating Universal Adversarial Perturbations for Quantum Classifiers. Work done by Gautham Govind Anil and Vishnu Vinod while at the University of British Columbia.

Paper

Title: Generating Universal Adversarial Perturbations for Quantum Classifiers
Authors: Gautham Govind Anil, Vishnu Vinod and Apurva Narayan
Conference: The 38th Annual AAAI Conference on Artificial Intelligence, February 20-27, 2024
Link to Paper (with supplementary material): Here AAAI Proceedings (Main Paper only): Here

Code Base

Folder structure of the Project Directory is given below

    ├── data            [datasets for XOR and Transverse Ising Model (TIM) classification tasks]
    │   ├── ising
    │   │   ├── 2       [2 qubit data]
    |   |   │   └── ...     (2 files: train and test)
    │   │   └── 4       [4 qubit data]
    |   |       └── ... (2 files: train and test)
    │   ├── xor
    │   │   └── ...     (2 files: train and test)
    │   └── dataset.py  [Code to generate datasets]
    ├── logs            [Directory to store all logs]
    │   └── ...         (empty)
    ├── results         [Directory to store all results]
    │   └── ...         (empty)
    ├── scripts         [Bash scripts for running different tasks (detailed later)]
    │   ├── test_cgen.sh
    │   ├── train_cgen.sh
    │   ├── train_qbim.sh
    │   ├── train_qgen.sh
    │   ├── train_target.sh
    │   └── train_unitary.sh
    ├── targets         [Trained Target models (classical and PQC-based classifiers)]
    │   └── ...         (18 files)
    ├── uaps            [Trained untargeted additive UAPs for testing (we provide UAPs with strengths 0.15 & 0.30)]
    │   └── ...         (30 files)
    ├── LICENSE
    ├── README.md
    ├── config.py
    ├── models.py
    ├── qmodels.py
    ├── requirements.txt
    ├── test_cgen.sh
    ├── train_cgen.sh
    ├── train_qbim.sh
    ├── train_qgen.sh
    ├── train_target.sh
    ├── train_unitary.sh
    └── utils.py

Supported Tasks

We support testing for the following classification tasks:

• MNIST/FMNIST - 2 class - $16\times16$
• MNIST/FMNIST - 4 class - $16\times16$
• MNIST/FMNIST - 2 class - $8\times8$
• TIM Dataset - 2 class

Kindly note that MNIST/FMNIST $8\times8$ is a special case designed for use only for testing the quantum generator in train_qgen.sh and train_qbim.sh.

Configuring the Task

The config.py contains $4$ different configurations each for a separate test setting. The default setting is Setting 4. Please set the correct setting for the task you test in the config.py file. Failure to do so will lead to erroneous results. The settings can be chosen by simply commenting out the unrequired setting and using the correct one as follows:

• Setting 1: Ising 2-class classification - 5 qubits, 1 ancillary qubit
• Setting 2: MNIST/FMNIST $8\times8$ - 2-class classification - 7 qubits, 1 ancillary qubit
• Setting 3: MNIST/FMNIST $16\times16$ - 2-class classification - 9 qubits, 1 ancillary qubit
• Setting 4: MNIST/FMNIST $16\times16$ - 4-class classification - 10 qubits, 2 ancillary qubits

Computational Requirements

This is necessary because quantum devices have to be initialized with a fixed number of wires. The code is structured in such a way. Kindly note that CUDA devices are required in order to read the pickled files. Please test on machines which have CUDA enabled devices and a RAM of atleast $20$ GB in order to allow flawless execution.

Python Scripts

Please run all python scripts from within the scripts/ folder

General code for running any of the scripts is given as bash script_file_name.sh

1. train_target.sh: Code to train target models (both classical and quantum)
2. train_cgen.sh: Code to train QuGAP-A for additive UAPs (control settings by changing)
3. test_cgen.sh: Code to test generated additive UAPs (RUN DIRECTLY TO SEE RESULTS)
4. train_unitary.sh: Code to train and test unitary simulation of QuGAP-U
5. train_qbim.sh: Code to train and test qBIM algorithm
6. train_qgen.sh: Code to train and test QuGAP-U algorithm

Further comments can be found inside each .sh file to aid in modifying, testing and playing around with the settings of each task.

Citation

If you use the work presented in the paper or the code hosted in this repository, please consider citing our paper at:

@article{
    Anil_Vinod_Narayan_2024, 
    title={Generating Universal Adversarial Perturbations for Quantum Classifiers}, 
    author={Anil, Gautham and Vinod, Vishnu and Narayan, Apurva}, 
    volume={38}, 
    url={https://ojs.aaai.org/index.php/AAAI/article/view/28963}, 
    DOI={10.1609/aaai.v38i10.28963}, 
    number={10}, 
    journal={Proceedings of the AAAI Conference on Artificial Intelligence}, 
    year={2024}, 
    month={Mar.}, 
    pages={10891-10899}
}