https://github.com/soskek/interval-bound-propagation-chainer

Sven Gowal et al., Scalable Verified Training for Provably Robust Image Classification, ICCV 2019
https://github.com/soskek/interval-bound-propagation-chainer

adversarial-attacks chainer deep-learning deepmind iccv2019 interval-bound-propagation neural-networks verification

Last synced: 2 months ago
JSON representation

Sven Gowal et al., Scalable Verified Training for Provably Robust Image Classification, ICCV 2019

• Host: GitHub
• URL: https://github.com/soskek/interval-bound-propagation-chainer
• Owner: soskek
• Created: 2019-09-23T05:42:33.000Z (about 6 years ago)
• Default Branch: master
• Last Pushed: 2019-11-17T15:36:37.000Z (almost 6 years ago)
• Last Synced: 2025-04-15T12:47:21.367Z (6 months ago)
• Topics: adversarial-attacks, chainer, deep-learning, deepmind, iccv2019, interval-bound-propagation, neural-networks, verification
• Language: Jupyter Notebook
• Homepage:
• Size: 246 KB
• Stars: 9
• Watchers: 3
• Forks: 2
• Open Issues: 0
• Metadata Files:
  • Readme: README.md

Awesome Lists containing this project

README

          
# Interval Bound Propagation

This is Chainer easy-to-follow implementation of Interval Bound Propagation. The MNIST experiment with small or medium models is implemented.

Paper: Sven Gowal, Krishnamurthy Dvijotham, Robert Stanforth, Rudy Bunel, Chongli Qin, Jonathan Uesato, Relja Arandjelovic, Timothy Mann, Pushmeet Kohli, [On the Effectiveness of Interval Bound Propagation for Training Verifiably Robust Models (Scalable Verified Training for Provably Robust Image Classification), ICCV 2019](https://arxiv.org/abs/1810.12715)

Authors' TensorFlow Code: [https://github.com/deepmind/interval-bound-propagation](https://github.com/deepmind/interval-bound-propagation)

```

python train_mnist.py -d 0 --model-class small

```

See `layers.py` or `models.py` for understanding the core of the algorithm.

## Visualize

![visualization of interval bound propagation CNN filters](sample_data/vis_files/visualize_comparison_ibp.png)

The left is by a baseline model while the right is by a IBP-trained model.

Each shows the activation feature map of `normal`, `upper`- and `lower`-bounds at each layer.

The image of `range` shows the `upper - normal` and `normal - lower` diffs with red and blue respectively.

The more red, the more looser the upper bound is. The more blue, the more looser the lower bound is.

We can see the IBP-trained model produces noise-robust features.

And, its logit for classification (`y`), on the bottom, is also robust and consistently predicts 7 as the label while the baseline fails.

Visualization notebook is `visualize_interval_bound_propagation.ipynb`.