Towards Analyzing Semantic Robustness of Deep Neural Networks (ECCV 2020)

Paper | Video | Tutorial.

Pytorch implementation of the paper awrded best paper at ECCV'20 Workshop on Adversarial Robustness in the Real World. The paper tries to address the robustness of Deep Neeural Networks, but not from pixel-level perturbation lense, rather from semantic lense in which the perturbation happens in the latent parameters that generate the image. This type of robustness is important for safety-critical applications like self-driving cars in which tolerance of error is very low and risk of failure is high.

Towards Analyzing Semantic Robustness of Deep Neural Networks
Abdullah Hamdi, Bernard Ghanem

Citation

If you find this useful for your research, please use the following.

@article{hamdi2019towards,
  title={Towards Analyzing Semantic Robustness of Deep Neural Networks},
  author={Hamdi, Abdullah and Ghanem, Bernard},
  journal={arXiv preprint arXiv:1904.04621},
  year={2019}
}

Examples of Results

• visualizing the Deep networks semantic profiles (2D) for 100 objects from 10 classes.

• Detecting robust regions of the networks with bounds-optimzing algorithms (2D).

Dataset

• We collect 100 3D shapes from 10 classes from ShapeNet. The details are found here

Prerequisites

• Linux
• Python 2 or 3
• NVIDIA GPU (11G memory or larger) + CUDA cuDNN

Getting Started

Installation

• install anaconda and then run the following commans

conda env create -f environment.yaml
source activate semantic
conda install -c anaconda cudatoolkit==9.0
pip install git+https://github.com/daniilidis-group/neural_renderer

• Clone this repo:

git clone https://github.com/ajhamdi/semantic-robustness
cd semantic-robustness

Simple Colab Tutorial with a toy example:

We provide a simple tutorial on a Colab notebook here and on Jupyter notebook here to test a toy example on some 3D objects and apply the bound optimzation algorithms for sample points in the 1D case directly on the cloud. The complete results obtained in the results directory are obtained as of the following sections

Running the experiments

1. Mapping the deep networks fully

you have to run map.py like the following:

python map.py --network resnet --gpu 0 --class_nb 1 --object_nb 0 --override

1. Finding the robust regions of the networks by the optimization

you have to run optim.py like the following:

python optim.py --network resnet --gpu 0 --class_nb 1 --object_nb 0 --override

The arguments descriptions are as follows:

• network : trained network type used in the experiments, choices=('incept', 'alexnet', 'vgg',"resnet") corresponding to (InceptionV3, ResNet50, VGG, AlexNet)
• gpu : the GPU number in which the exp perfoprmed
• class_nb : number of the class used for the experiment (0-9) of the dataset
• object_nb : number of the 3D object used for the experiment (0-9) of the dataset
• override : a flag to override exisisting results

saving the results:

The checkpoint directory contains the results as dictionaries and is arranged as follws :

├── ...
   ├── checkpont                    # contaong the optimzatations traces of all experiments 
   │   ├── NETWORK_NAME           # the network
   │   |    ├── CLASS_NUMBER           # class number (0-9) whcih is part of the 10 3D classes above and also part of ImageNet Classes 
                ├── OBJECT_NUMBER    #number of the object (0-9) from the 10 objects in that specific class 
   │   └── ...               
   └── ...

The results from optim.py will be saved as dictionaries to the directory : ./checkpoint/NETWORK_NAME/CLASS_NUMBER/OBJECT_NUMBER/optim.pt. and the mapping results from map.py will be wsaved in ./checkpoint/NETWORK_NAME/CLASS_NUMBER/OBJECT_NUMBER/map.pt

The map.pt dictionary contains the following:

map_dict = torch.load("checkpoint/ResNet50/0/0/map.pt")
map_dict
{
"xx": # azimuth parameters list  
"yy" : # elevation parameters list
"z" : # the network confidence 
}

The optim.pt dictionary contains the optimization results as follows:

optim_dict = torch.load("checkpoint/ResNet50/0/0/optim.pt")
optim_dict
{
"initial_point": # the initial point used in the optimization 
"network_name" : # the network name used in the evaluation ["ResNet50" , "Inceptionv3", "VGG", "AlexNet"]
"class_nb" : # the class nb [0-9] used in the experiments
"OIR_W" OR "OIR_B" OR "naive" = #  the type of optimization method used 
   {"optim_trace" : # the trace of the bounds through the the optimization (list of traces for every initial point)
    "loss_trace" : #the trace of the loss throug the optimization (list of traces for every initial point)
    "regions": #last region converged to (list of ndintervall class  for every initial point) 
   }
}

Testing with your own 3D dataset and networks

• TODO NEXT

Acknowledgments

This work is suuported by King Abdullah University of Science and Technology (KAUST).The code borrows heavily from neural mesh renderer.