Skip to main content

Differentiable abstract domain implementations for neural network reasoning on PyTorch

Project description


DiffAbs is a PyTorch implementation of multiple abstract domains that can be used in certifying or reasoning neural networks. Implemented purely using PyTorch, it is differentiable and supports GPU by default, thus amenable for safety/robustness driven training on abstract domains.

Currently, the following abstract domains are implemented:

Domain notes

DeepPoly ReLU heuristics:

  • A variant of the original DeepPoly domain is implemented where the ReLU approximation is not heuristically choosing between two choices (either picking y = x or y = 0 as the new upper bound). Right now it is fixed to choosing y = 0, because there was Galois connection violation observed if this heuristic is enabled. Basically, it is observed in experiment that a smaller abstraction may unexpectedly incur larger safety distance than its containing larger abstraction.

Supported systems

Although it is currently tested on Mac OS X 10.15 and Ubuntu 16.04 with Python 3.7 and PyTorch 1.5, it should generalize to other platforms and older PyTorch (perhaps ≥ v1.0) smoothly.

However, Python ≤ 3.6 may be incompatible. Because type annotations are specified everywhere and the type annotation of self class is only supported by __future__.annotations in Python 3.7. If using Python 3.6, this needs to use 'type string' instead.


In your virtual environment, either install directly from this repository by

git clone
cd DiffAbs
pip install -e .

or directly install from PyPI:

pip install diffabs


Test cases for individual abstract domains are under the tests/ directory and can be run using command


Note that torchvision is needed to run the tests for conv/maxpool layers.


The project is available open source under the terms of MIT License.

Project details

Release history Release notifications | RSS feed

This version


Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

diffabs-0.1.tar.gz (21.1 kB view hashes)

Uploaded source

Built Distribution

diffabs-0.1-py3-none-any.whl (23.6 kB view hashes)

Uploaded py3

Supported by

AWS AWS Cloud computing and Security Sponsor Datadog Datadog Monitoring Fastly Fastly CDN Google Google Download Analytics Microsoft Microsoft PSF Sponsor Pingdom Pingdom Monitoring Sentry Sentry Error logging StatusPage StatusPage Status page