Veritex is an object-oriented software tool programmed in Python for verifying and repairing neural networks. It takes in two inputs, the network model and safety properties. Veritex supports the standardized format ONNX and PyTorch for the network and the unified format Vnnnlib for the safety property. With the network model and its safety properties, Veritex can:
- compute the exact or over-approximated output reachable domain and also the entire unsafe input space if exists,
- plot 2 or 3-dimensional reachable domains,
- produce a provable safe network in ONNX or PyTorch format when the repair option is enabled.
Figure: An overview of Veritex architecture.
Clone this repository to your local machine, which we assume has 32GB RAM or more, and where the artifact is executable through Docker on most common host operating systems (Linux/Ubuntu and Windows 11 have been tested).
git clone https://github.com/Shaddadi/veritex.git
cd veritex
-
Assuming Docker is installed, build the Docker image from the dockerfile (ignore sudo if using a Windows host in these steps).
sudo docker build . -t veritex_image
-
Create and start the docker container.
sudo docker run --name veritex --rm -it veritex_image bash
This tool has been confirmed to work and tested with only Python3.7.
-
Install veritex pkg with pip.
python -m pip install .
-
Set path to /veritex under this repository.
export PYTHONPATH='<YOUR_REPO_PATH>/veritex' export OPENBLAS_NUM_THREADS=1 export OMP_NUM_THREADS=1
This tool has been confirmed to work and tested with only Python3.7.
-
Install required python packages.
python -m pip install -e .
-
Set path to /veritex under this repository.
export PYTHONPATH='<YOUR_REPO_PATH>/veritex' export OPENBLAS_NUM_THREADS=1 export OMP_NUM_THREADS=1
Linux systems are suggested. This artifact aims to reproduce results in the HSCC'23 tool paper, including Figure 2&3&4 and Table 2&3. Results are stored in 'veritex/artifact/HSCC23/results'. There are two versions for the artifact evaluation. The difference between these two versions is that the first one does not include the repair of two neural networks which consumes a large amount of memory and time.
Caution: Reachable domains of networks in Figure 3&4 may be slightly different from the ones in the paper because each run of the repair method can not guarantee to produce the exact same safe network.
Caution: On Windows hosts, users who encounter the error '\r command not found' when executing the artifact, please run the following commands before the shell script. The update may not be required, but the dos2unix tool can be used to address the line endings if this error arises. Note this should be done in the HSCC23 directory, where these reproduce_resultsX.sh scripts reside.
apt-get update
apt-get install dos2unix
dos2unix reproduce_results1.sh
dos2unix reproduce_results2.sh
-
The first version reproduces the results in the paper except for two hard instances (~170 mins), including
- safety verification of all instances (data generation for Figure 2) (~2 mins),
- repair of 33/35 unsafe instances (data generation for Figure 3 and most of results in Table 2&3) (~40 mins),
- repair of an unsafe DNN agent (data generation for Figure 4) (~6 mins),
- implementation of the related work ART for the repair comparison (~90 mins),
- generation of figures and tables (~40 mins, majority of the time is spent on the plot of reachable domains).
This version requires at least 32 GB memory.
cd artifact/HSCC23 bash reproduce_results1.sh
-
The second version reproduces all the results in the paper (~410 mins), including
- safety verification of all instances (data generation for Figure 2) (~2 mins),
- repair of all 35/35 unsafe instances (data generation for Figure 3 and Table 2&3) (~280 mins),
- repair of an unsafe DNN agent (data generation for Figure 4) (~6 mins),
- implementation of the related work ART for the repair comparison (~90 mins),
- generation of figures and tables (~40 mins, majority of the time is spent on the plot of reachable domains).
The hardware requirement for second version is AWS, CPU: r5.12xlarge, 48vCPUs, 384 GB memory, no GPU.
cd artifact/HSCC23 bash reproduce_results2.sh
-
Export results from docker to host.
sudo docker cp veritex:/veritex/artifact/HSCC23/results/. <PATH_TO_YOUR_HOST>
This section describes how to run additional demonstrations and experiments beyond those reproduced above.
This demo includes the computation of the exact output reachable domain of a neural network using our reachability analysis method. It also includes the computation of its exact unsafe input space that leads to safety violations in the output using our Backtracking algorithm. The neural network consists of 3 inputs, 2 outputs, and 8 layers with each having 7 neurons. Results will be saved in '/images'.
cd examples/Demo
python main_demo.py
Figure: Demo for our reachability analysis algorithms. Given an input domain (the blue box), our algorithms compute the exact output reachable domain and also the exact unsafe input subspace that leads to safety violation in the output domain.
Reuse: compute the exact output rachable domain and the unsafe input subspace of a DNN
- Create a DNN object using FFNN and enable 'exact_outputd' and 'unsafe_inputd'
- Create a safety property by specifying the lower and upper bounds of the input domain and the set representation'FVIM' or 'Flattice'
- Set up and run the parallel framework using SharedState and Worker
- Collect results and plot using reachplot
This section describes how to verify a new neural network not already analyzed within the Veritex examples, illustrated as well for an ACASXu example.
cd examples/ACASXu/verify
python verify_all_instances.py
Reuse: verify a neural network on multiple safety properties using verify
python veritex/methods/verify.py \
--property 'path_to_property1' ... 'path_to_propertyn' \
--property_name 'name1' ... 'namen' \
--network_path 'path_to_model' \
--network_name 'model_name'
ACASXu Example. Results will be logged.
cd examples/ACASXu/verify
python ../../../veritex/methods/verify.py \
--property '../nets/prop_1.vnnlib' '../nets/prop_2.vnnlib' '../nets/prop_3.vnnlib' '../nets/prop_4.vnnlib' \
--property_name 'prop_1' 'prop_2' 'prop_3' 'prop_4' \
--network_path '../nets/ACASXU_run2a_1_1_batch_2000.onnx' \
--network_name 'N11'
This section describes how to repair a new neural network not already analyzed within the Veritex examples.
cd examples/ACASXu/repair
python repair_nnets.py
Reuse: repair a neural network on multiple safety properties
- Build a repair object using REPAIR
- Select parameters for the retraining process
- Execute the repair function repair_model_classification or repair_model_regular
Visualize the output reachable domain using reachplot
python veritex/methods/reachplot.py \
--property 'path_to_property1'...'path_to_propertyn' \
--network_path 'path_to_model' \
--dims x x \
--savename 'xxx'
Example. Results will be logged into '/figures':
cd examples/ACASXu/repair
python ../../../veritex/methods/reachplot.py \
--property '../nets/prop_3.vnnlib' '../nets/prop_4.vnnlib' \
--network_path '../nets/ACASXU_run2a_2_1_batch_2000.onnx' \
--dims 0 1 \
--savename 'figures/reachable_domain_property_3,4_dims0_1'
python ../../../veritex/methods/reachplot.py \
--property '../nets/prop_3.vnnlib' '../nets/prop_4.vnnlib' \
--network_path '../nets/ACASXU_run2a_2_1_batch_2000.onnx' \
--dims 0 2 \
--savename 'figures/reachable_domain_property_3,4_dims0_2'
Figure: Output reachable domains of Network21 on Properties 3 and 4. They are projected on (y0,y1) and (y0, y2).
Over approximate reachable domains of networks Demo. Results will be logged into '/figures'.
cd examples/Demo
python demo_vzono.py
Figure: Over approximated reachable domains of networks with ReLU, Tanh, Sigmoid, respectively.
Reuse: over approximate output reachable domain of a DNN with Vzono
- Create a DNN object using FFNN
- Create a safety property by specifying the lower and upper bounds of the input domain and the set representation Vzono
- Set up and run the parallel framework using SharedState and Worker
- Collect results and plot using reachplot
Fast reachability analysis of pixelwise perturbation of CNNs using Method. Results will be logged into '/figures'.
examples/CIFAR10
python main_cifar10.py
Figure: Fast Reachability analysis of CNNs with input pixels under perturbation. Blue area represents the reachable domain. Black dots represent simultations.