L_infinity-distance Net is a theoretically principled neural network that is inherently robust to L_infinity-norm perturbations. It was first proposed in ICML 2021 and implemented here. L_infinity-distance Net consistently achieves state-of-the-art certified L_infinity robustness on multiple datasets.
This github repository gives an upgrade of L_infinity-distance Net. The main features are described as follows.
We significantly speed up the calculation of L_p-distance operation with a better CUDA implementation, both for forward propagation and backward propagation. It therefore makes the training process considerably faster.
Our implementation adopts different parallel algorithms for the three cases: when p is a small integer (e.g. 8), when p is infinity, and when p is a general floating-point number. The extent of acceleration in each case is shown in the table below (tested on a NVIDIA RTX 3090 GPU).
Table: Comparison of the wall-clock time (seconds) of a training iteration (forward+backward propagation).
p | Previous implementation | Our implementation | Speed up |
---|---|---|---|
general | 0.2927 | 0.1866 | 56% |
small integer | 0.2927 | 0.1133 | 160% |
infinity | 0.09433 | 0.07961 | 19% |
With the accelerated CUDA implementation, training an L_infinity-distance net (800 epochs) only takes about 3.7 hours.
We design an improved training strategy that substantially improves the certified robust accuracy on multiple datasets. Results are listed in the table below. See our latest paper (titled "Boosting the Certified Robustness of L-infinity Distance Nets") for details of the training strategy.
Dataset | eps | Clean Acc | PGD-100 Acc | Certified Acc |
---|---|---|---|---|
MNIST | 0.1 | 98.93 | 98.03 | 97.95 |
MNIST | 0.3 | 98.56 | 94.73 | 93.20 |
CIFAR-10 | 2/255 | 60.61 | 54.28 | 54.12 |
CIFAR-10 | 8/255 | 54.30 | 41.84 | 40.06 |
CIFAR-10 | 16/255 | 48.50 | 32.73 | 29.04 |
To facilitate future research, we provide an interface containing the major components of this repository in the core
folder. This library is portable and you can simply copy this folder to your work space and use it outside the box. We list the functionalities of the core library:
- A function
norm_dist(x, w, p)
that calculates the L_p distance between the input tensorx
and the parameter tensorw
with CUDA acceleration (in core/norm_dist.py). - A base class
NormDistBase
that wraps the above function into a model layer suitable for pytorchtorch.nn.Module
. Then subclassesNormDist
andNormDistConv
specialize the fully-connected layer and convolution layer (in core/modules/norm_dist.py). - The library also supports Interval Bound Propagation (IBP). We provide IBP layer for the linear transformation, convolution, and commonly used activation functions (in core/modules/basic_modules.py) as well as the L-infinity distance function. They can be used to build the top MLP to form a composite architecture (L_infinity-distance net + MLP).
- Pytorch 1.8.0 (or a later version)
- Tensorboard (optional)
After cloning this repo into your computer, first run the following command to install the CUDA extension.
python setup.py install --user
We provide complete training scripts to reproduce the results in our latest paper. These scripts are in the command
folder.
For example, to reproduce the result of CIFAR-10 with perturbation eps=8/255, simply run
bash command/cifar_0.03137.sh
Reproducing baseline results in ICML 2021
For example, to reproduce the baseline results of L_infinity-distance net on CIFAR-10 dataset with perturbation eps=8/255, run the following command
python main.py --dataset CIFAR10 --model 'MLPModel(depth=6,width=5120,identity_val=10.0,scalar=False)' --loss 'hinge' --p-start 8 --p-end 1000 --epochs 0,0,100,750,800 --eps-test 0.03137 --eps-train 0.1569 -b 512 --lr 0.02 --gpu 0 -p 200 --seed 2021 --visualize
To reproduce the baseline results of L_infinity-distance net + MLP on CIFAR-10 dataset with perturbation eps=8/255, run the following command
python main.py --dataset CIFAR10 --model 'HybridModel(depth=7,width=5120,identity_val=10.0,hidden=512)' --loss 'crossentropy' --p-start 8 --p-end 1000 --epochs 0,100,100,750,800 --eps-test 0.03137 --eps-train 0.03451 -b 512 --lr 0.02 --gpu 0 -p 200 --seed 2021 --visualize
We also support multi-GPU training using distributed data parallel. By default the code will use all available GPUs for training. To use a single GPU, add the following parameter --gpu GPU_ID
where GPU_ID
is the GPU ID. You can also specify --world-size
, --rank
and --dist-url
for advanced multi-GPU training.
The model is automatically saved when the training procedure finishes. Use --checkpoint model_file_name.pth
to load a specified model before training. You can use --start-epoch NUM_EPOCHS
to skip training and only test the model's performance for a pretrained model, where NUM_EPOCHS
is the number of epochs in total.
By default the code will generate five files named train.log
, test.log
, train_inf.log
, test_inf.log
and log.txt
which contain all training logs. If you want to further display training curves, you can add the parameter --visualize
to show these curves using Tensorboard.
We also provide pretrained models with SOTA certified robust accuracy. These models can be downloaded here. The models are compressed into 7z file format. To use these models, follow the Saving and Loading instruction above.
Please contact zhangbohang@pku.edu.cn if you have any question on our paper or the codes. Enjoy!
@inproceedings{zhang2021towards,
title={Towards Certifying L-infinity Robustness using Neural Networks with L-inf-dist Neurons},
author={Zhang, Bohang and Cai, Tianle and Lu, Zhou and He, Di and Wang, Liwei},
booktitle={International Conference on Machine Learning},
pages={12368--12379},
year={2021},
organization={PMLR}
}
@inproceedings{zhang2022boosting,
title={Boosting the Certified Robustness of L-infinity Distance Nets},
author={Bohang Zhang and Du Jiang and Di He and Liwei Wang},
booktitle={International Conference on Learning Representations},
year={2022},
}