dnnf 0.1.3

dnnf - dnn property falsification

Reducing DNN Properties to Enable Falsification with Adversarial Attacks

This repo accompanies the paper Reducing DNN Properties to Enable Falsification with Adversarial Attacks, and provides a tool for running falsification methods such as adversarial attacks on DNN property specifications specified using the DNNP language of DNNV. For an overview of our paper, check out our video presentation.

Additional documentation can be found on Read the Docs.

Install

We provide instructions for installing DNNF with pip, installing DNNF from source, as well as for building and running a docker image.

Pip Install

DNNF can be installed using pip by running:

  $ pip install dnnf

This will install the last version uploaded to PyPI. To install the most recent changes from GitHub, run:

  $ pip install git+https://github.com/dlshriver/DNNF.git@main

To install the cleverhans or foolbox backends, run the above command with the option --install-option="--extras-require=cleverhans,foolbox" included.

Note: installation with pip will not install the TensorFuzz falsification backend. Currently this backend is only available through manual installation or the provided docker image.

Source Install

The required dependencies for installation from source are:

• git
• virtualenv
• python3.7
• python3.7-dev
• python2.7

Please ensure that these dependencies are installed prior to running the rest of the installation script. For example, on a fresh Ubuntu 20.04 system, the dependencies can be installed using apt as follows:

  $ sudo add-apt-repository ppa:deadsnakes/ppa
  $ sudo apt-get update
  $ sudo apt-get install python3.7
  $ sudo apt-get install python3.7-dev
  $ sudo apt-get install python2.7
  $ sudo apt-get install virtualenv
  $ sudo apt-get install git

To install DNNF in the local directory with all available backend falsification methods, download this repo and run the provided installation script:

  $ ./install.sh --include-cleverhans --include-foolbox --include-tensorfuzz

To see additional installation options, use the -h option.

We have successfully tested this installation procedure on machines running Ubuntu 20.04 and CentOS 7.

Docker Install

We provide a pre-built docker image containing DNNF, available on Docker Hub. To use this image, run the following:

  $ docker pull dlshriver/dnnf
  $ docker run -it dlshriver/dnnf
  (.venv) dnnf@hostname:~$ dnnf -h

To build a docker image with the latest changes to DNNF, run:

  $ docker build . -t dlshriver/dnnf
  $ docker run -it dlshriver/dnnf
  (.venv) dnnf@hostname:~$ dnnf -h

Execution

To execute DNNF, first activate the virtual environment with:

  $ . .venv/bin/activate

This is only required if DNNF was installed manually. The virtual environment should open automatically if using the docker image.

The DNNF tool can then be run as follows:

  $ dnnf PROPERTY --network NAME PATH

Where PROPERTY is the path to the property specification, NAME is the name of the network used in the property specification (typically N), and PATH is the path to a DNN model in the ONNX format.

To see additional options, run:

  $ dnnf -h

Running on the Benchmarks

We provide the property and network benchmarks used in our evaluation here.

To execute DNNF on a problem in one of the benchmarks, first navigate to the desired benchmark directory in artifacts (i.e., acas_benchmark, neurifydave_benchmark, or ghpr_benchmark). Then run DNNF as specified above. For example, to run DNNF with the Projected Gradient Descent adversarial attack from cleverhans on an ACAS property and network, run:

  $ cd artifacts/acas_benchmark
  $ dnnf properties/property_2.py --network N onnx/N_3_1.onnx --backend cleverhans.ProjectedGradientDescent

Which will produce output similar to:

  Falsifying: Forall(x0, (((x0 <= [[ 0.68 0.5  0.5  0.5 -0.45]]) & ([[ 0.6 -0.5 -0.5  0.45 -0.5 ]] <= x0)) ==> (numpy.argmax(N(x0)) != 0)))

  dnnf
    result: sat
    time: 2.6067

The available backends for falsification are:

• cleverhans.LBFGS, which also requires setting parameters --set cleverhans.LBFGS y_target "[[-1.0, 0.0]]"
• cleverhans.BasicIterativeMethod
• cleverhans.FastGradientMethod
• cleverhans.DeepFool, which also requires setting parameters --set cleverhans.DeepFool nb_candidate 2
• cleverhans.ProjectedGradientDescent
• tensorfuzz

If a property uses parameters, then the parameter value can be set using --prop.PARAMETER=VALUE, e.g., --prop.epsilon=1, similar to DNNV.

Acknowledgements

This material is based in part upon work supported by the National Science Foundation under grant number 1900676 and 2019239.