This repository contains the code for re-producing the experiments of the papers:
- Quantum Robustness Verification: A Hybrid Quantum-Classical Neural Network Certification Algorithm
- Efficient MILP Decomposition in Quantum Computing for ReLU Network Robustness
presented at IEEE Quantum Week - IEEE International Conference on Quantum Computing and Engineering (QCE) in 2022 and 2023, respectively.
Robustness certification of neural networks can be modelled as Mixed-Integer Linear Program (MILP).
HQ-CRAN uses two decomposition methods (Benders & Dantzig-Wolfe) to split the original MILP problem into a linear program (LP) and quadratic unconstrained binary optimization (QUBO), solved by classical and quantum algorithm, respectively.
We evaluate HQ-CRAN against one complete verifier:
and two incomplete verfiers:
In terms of certified accuracy, HQ-CRAN-BD performs similarly to 𝛽-CROWN. On the contrary, the average runtime is ∼2 orders of magnitude slower than incomplete methods and ∼1 order than 𝛽-CROWN.
In the context of HQ-CRAN-DW, the number of certified samples is similar than GPUPoly and PRIMA for
Create a minimal environment with only Python installed in it.
conda create -y -n hq-cran python=3.8Activate your new environment.
conda activate hq-cranNext, install the requirements.
pip install -r requirementsActivate your environment.
conda activate qc-certIn this folder run
python hqcran/. --config config_example.yamlonnxpath: str = None # ONNX network path
netname: str = None # network name
torchpath: str = None # torch network path
dataset: str = None # dataset path (CSV file)
output: str = None # output folder (python dictionary)
solver: str = None # solver type: cplex, ibm_simulator, ibm_cloud, dwave_hybrid, dwave_qpu, dwave_simulator
epsilon: float = 0.001 # adversarial power for the l-infinity norm
weight_t: int = 1 # penalization value for the quadratic term
weight_p: float = 0.01 # weights to properly adjust the approximation of p
weight_x: float = 0.01 # weights to properly adjust the approximation of x
sub_bound: int = 5 # maximum value for the sub bounds (alpha, beta)
objectives_gap: float = 0.1 # Gap between the master and sub objective (xi)
start: int = None # starting test
end: int = None # ending test
steps: int = 100 # maximum number of steps for the hybrid algorithm
classic: bool = True # pure classic solver to True if you want to have a comparison with classic solvers
dwave_reads: int = 100 # number of reads of the D-Wave QPU system
threads: int = 32 # maximum number of Threads
verbose: bool = False # Set debug to True if you want to visualize every output
real_var: bool = False # real variables for debugging the benders decomposition
qubo: bool = True # QUBO formulation to True (False works only with real variables)
magnanti: bool (Default True) # generation of Magnanti & Wong cuts
hamming: bool = False # strenght factor for the Hamming distance from the previous solution
max_cuts: int = np.inf # maximum number of cuts in the set
gurobi: bool = False # Gurobi optimizer instead of CPLEX
decomposition: str = None # decomposition strategy: can be either benders or dantzig-wolfe
crown_ibp: bool = True # verify sample with initial bound from CROWN IBPTo obtain a CPLEX academic license follow Get the no-cost academic edition at this link CPLEX OPTIMIZATION STUDIO. Go on software and select ILOG CPLEX OPTIMIZATION STUDIO and download. Create an account and follow the instructions. Once you get the linux bash script ILOG_COS_20.10_LINUX_X86_64.bin run it and define your local installation path /home/yourname/ibm/ILOG/CPLEX_Studio201. After that you can move inside the python folder (/home/yourname/ibm/ILOG/CPLEX_Studio201/python) and install the api with python setup.py install
Create an account on D-Wave Cloud and follow t 6B48 he instructions
This software was solely developed for and published as part of the publication cited above. It will neither be maintained nor monitored in any way.
Please cite our paper if you use HQ-CRAN in your research.
- Quantum Robustness Verification: A Hybrid Quantum-Classical Neural Network Certification Algorithm
@INPROCEEDINGS{9951230, author={Franco, Nicola and Wollschläger, Tom and Gao, Nicholas and Lorenz, Jeanette Miriam and Günnemann, Stephan}, booktitle={2022 IEEE International Conference on Quantum Computing and Engineering (QCE)}, title={Quantum Robustness Verification: A Hybrid Quantum-Classical Neural Network Certification Algorithm}, year={2022}, volume={}, number={}, pages={142-153}, doi={10.1109/QCE53715.2022.00033}}
- Efficient MILP Decomposition in Quantum Computing for ReLU Network Robustness
@INPROCEEDINGS{10313804, author={Franco, Nicola and Wollschläger, Tom and Poggel, Benedikt and Günnemann, Stephan and Lorenz, Jeanette Miriam}, booktitle={2023 IEEE International Conference on Quantum Computing and Engineering (QCE)}, title={Efficient MILP Decomposition in Quantum Computing for ReLU Network Robustness}, year={2023}, volume={01}, number={}, pages={524-534}, doi={10.1109/QCE57702.2023.00066}}