airbus/decomon

Decomon: Automatic Certified Perturbation Analysis of Neural Networks

Introduction

What is decomon? decomon is a library that allows the derivation of upper and lower bounds for the predictions of a Tensorflow/Keras neural network with perturbed inputs. In the current release, these bounds are represented as affine functions with respect to some variable under perturbation.

Previous works that tackled certified robustness with backward propagation relied on forward upper and lower bounds. In decomon, we explored various ways to tighten forward upper and lower bounds, while remaining backpropagation-compatible thanks to symbolic optimization.

Our algorithm improves existing forward linear relaxation algorithms for general Keras-based neural networks without manual derivation. Our implementation is also automatically differentiable. So far we support interval bound propagation, forward mode perturbation, backward mode perturbation as well as hybrid approaches.

decomon is compatible with a wider range of perturbation: boxes, $L_{\inf, 1, 2}$ norms or general convex sets described by their vertices.

We believe that decomon is a complementary tool to existing libraries for the certification of neural networks.

Since we rely on Tensorflow and not Pytorch, we are opening up the possibility for a new community to formally assess the robustness of their networks, without worrying about the technicality of the implementation. In this way, we hope to promote the formal certification of neural networks into safety critical systems.

Installation

Quick version:

pip install decomon

For more details, see the online documentation.

Quick start

You can see how to get certified lower and upper bounds for a basic Keras neural network in the Getting started section of the online documentation.

Documentation

The latest documentation is available online.

Examples

Some educational notebooks are available in tutorials/ folder. Links to launch them online with colab or binder are provided in the Tutorials section of the online documentation.

Contributing

We welcome any contribution. See more about how to contribute in the online documentation.

decomon's Issues

Tutorial 2 needs pandas library

tutorials/tutorial2_noise_sensor.ipynb need pandas which is not included natively when installing the lib with poetry. However, It is ok to not put all notebook dependencies into the library dependencies

Add compute_output_shape() for backward layers

Add a test for BackwardBatchNormalization

wrapper_with_tuning: non conversion to DecomonModel

Missing "poetry install" in Installation Instructions

After poetry build command, you need to also call poetry install, cheers

Import error ortools

ortools is not required in the main version of decomon and is not captured as a dependency
action:
either put a conditional import
or add it in the dependency list

Enable choice of slope in convert/clone

The choice of slope used when linearizing the activation functions is not available for high-level functions from wrapper.py or models/convert.py. This would be nice to allow this.
For now the default value V_SLOPE is used.

`utils_conv.get_toeplitz_channels_last()` assume a squared kernel

When extracting image patches, a squared filter size based on conv_layer.kernel.shape[0] is used.
We could perhaps use the actual kernel size, potentially rectangular: conv_layer.kernel_size (which is a 2-length tuple after Conv2D initialization).