CAISAR

CAISAR (Characterizing AI Safety And Robustness) is a platform under active development at CEA LIST, aiming to provide a wide range of features to characterize the safety and robustness of artificial intelligence based software.

Installation

No binaries are provided at the moment. Installation must be done by directly compiling the source.

Please note: CAISAR requires the OCaml package manager opam, v2.1 or higher, which is typically avaible in all major GNU/Linux distributions.

To build CAISAR, do the following:

$ git clone https://git.frama-c.com/pub/caisar
$ cd caisar
$ opam switch create --yes --no-install . 4.13.1
$ opam install . --deps-only --with-test --yes
$ make

To run the tests:

$ make test

Usage

To start using CAISAR, please run the command:

$ dune exec -- caisar --help

Property verification

CAISAR can be used to verify properties on neural networks and support-vector machines (SVM).

The prototype command is:

$ dune exec -- caisar verify --prover=PROVER FILE

FILE defines the property to verify, and it must be written in the WhyML language. Examples of WhyML files (.mlw) can be found in the tests folder.

External provers detection

CAISAR relies on external provers to work. These must be installed first, then CAISAR must be instructed to point to their location. To do so, the path to the prover executables should appear in the environment variable PATH.

Run the following command to confirm that CAISAR detects the installed provers:

$ dune exec -- caisar config --detect

The following are the provers for which a support is provided in CAISAR:

• PyRAT (to be released)
• Marabou
• SAVer

Under active development is the support for the SMT-LIB which is used by many satisfiability modulo theories solvers (e.g. Alt-Ergo, Z3, Colibri, etc.).

Advanced usage

How to add a solver

Make sure the solver is installed in your system. Typically, the path to its executable should appear in the environment variable PATH. Then,

1. Create a solver.drv in config/drivers/. A driver is a series of WhyML modules describing the theories the solver is able to understand as provided by Why3. Directives for letting Why3 interpret the solver outputs should also be provided here.

2. Add a new record in config/caisar-detection-data.conf. The name of the solver executable should be provided , as well as a command-line template that Why3 will use for executing the solver. Such a template may specify several Why3 built-in identifiers:

   • %e stands for the executable
   • %f stands for a file to pass to the executable

Other custom identifiers have been added: %{nnet-onnx} and %{svm}. These identifiers are used for providing the solver with potential {nnet, onnx} and svm model filenames, respectively.

3. Write a Why3 printer. The solver should be recognized by CAISAR by now. However, a printer for the solver may be needed for transforming Why3 specifications into something the solver can understand. Printers should be placed in src/printers/.