Perturbation-Constrained Flow Attack (PCFA)

This repository contains the source code for our ECCV 2022 paper A Perturbation-Constrained Adversarial Attack for Evaluating the Robustness of Optical Flow by J. Schmalfuss, P. Scholze and A. Bruhn. If you find this work useful, please cite us as

@InProceedings{Schmalfuss2022PCFA,
title = {A Perturbation-Constrained Adversarial Attack for Evaluating the Robustness of Optical Flow},
author = {Schmalfuss, Jenny and Scholze, Philipp and Bruhn, Andrés},
year = {2022},
booktitle = {Proc. European Conference on Computer Vision (ECCV)},
publisher = {Springer},
pages = {183--200}
}

Abstract

Recent optical flow methods are almost exclusively judged in terms of accuracy, while their robustness is often neglected. Although adversarial attacks offer a useful tool to perform such an analysis, current attacks on optical flow methods focus on real-world attacking scenarios rather than a worst case robustness assessment. Hence, in this work, we propose a novel adversarial attack - the Perturbation-Constrained Flow Attack (PCFA) – that emphasizes destructivity over applicability as a real-world attack. PCFA is a global attack that optimizes adversarial perturbations to shift the predicted flow towards a specified target flow, while keeping the L2 norm of the perturbation below a chosen bound. Our experiments demonstrate PCFA’s applicability in white- and black-box settings, and show it finds stronger adversarial samples than previous attacks. Based on these strong samples, we provide the first joint ranking of optical flow methods considering both prediction quality and adversarial robustness, which reveals state-of-the-art methods to be particularly vulnerable.

Also refer to our preprint for details on the method.

Numeric values for Quality-Robustness analysis (Fig. 7)

Method	FlowNet2	SpyNet	PWCNet	RAFT	GMA
Adversarial Robustness KITTI15	20.84	9.61	20.26	27.75	25.48
Adversarial Robustness Sintel	17.67	12.24	16.53	19.29	22.91

Initial setup

Setup virtual environment

python3 -m venv pcfa
source pcfa/bin/activate

Install required packages:

Change into scripts folder and execute the script which installs all required packages via pip. As each package is installed succesively, you can debug errors for specific packages later.

cd scripts
bash install_packages.sh

Spatial Correlation Sampler

If the installation of the spatial-correlation-sampler works and you have a cuda capable machine, open helper_functions/config_paths.py and make sure to set the variable "correlationSamplerOnlyCPU": to False. This will speed up computations when using PWCNet.

If the spatial-correlation-sampler does not install run the following script to install a cpu-only version:

cd scripts
bash install_scs_cpu.sh

Loading Flow Models

Download the weights for a specific model by changing into the scripts/ directory and executing the bash script for a specific model:

cd scripts
./load_[model]_weights.sh

Here [model] should be replaced by one of the following options:

[ all | raft | gma | spynet | pwcnet | flownet2]

Note: the load_model scripts remove .git files, which are often write-protected and then require an additional confirmation on removal. To automate this process, consider to execute instead

yes | ./load_[model]_weights.sh

Compiling Cuda Extensions for FlowNet2

Please refer to the pytorch documentation how to compile the channelnorm, correlation and resample2d extensions. If all else fails, go to the extension folders /models/FlowNet/{channelnorm,correlation,resample2d}_package, manually execute

python3 setup.py install

and potentially replace cxx_args = ['-std=c++11'] by cxx_args = ['-std=c++14'], and the list of nvcc_args by nvcc_args = [] in every setup.py file. If manually compiling worked, you may need to add the paths to the respective .egg files in the {channelnorm,correlation,resample2d}.py files, e.g. for channelnorm via

sys.path.append("/lib/pythonX.X/site-packages/channelnorm_cuda-0.0.0-py3.6-linux-x86_64.egg")
import channelnorm_cuda

The site-packages folder location varies depending on your operation system and python version.

Datasets

For training and evaluation in PCFA we use the Sintel and KITTI 2015 dataset. Set correct paths to datasets (after downloading them) in helper_functions/config_paths.py

Code Usage

Training Perturbations with PCFA

To train perturbations with PCFA, execute

python3 attack_PCFA.py --net=[SpyNet,PWCNet,RAFT,GMA,FlowNet2]

By default, this trains disjoint image-specific perturbation with L2 bound 0.005 and a zero-flow target on KITTI15 evaluation, and saves output after every image. All available argument options are displayed via

python3 attack_PCFA.py --help

The following arguments are useful to reproduce the paper results:

--net=[SpyNet,PWCNet,RAFT,GMA,FlowNet2] | The network for which to execute
--delta_bound=0.005             | The average L2 bound that is used to train the perturbation, std=0.005

--joint_perturbation            | If this argument is passed, a joint perturbation is trained.
--universal_perturbation        | If this argument is passed, a universal perturbation is trained.

--target=[zero,neg_flow,custom] | Allows to set a target. If a custom target is desired, specify the path to a flow-containing file via --custom_target_path.
--loss=[aee,mse,cosim]          | Allows to specify the loss function
--steps=20                      | The number of LBFGS steps per image. Consider a lower number for --universal_perturbations (e.g. 1).
--boxconstraint=[clipping,change_of_variables]   | Allows to specify the box constraint

By default, universal perturbations are trained for --epochs=25 with a batch-size of --batch_size=4; both values are treated as 1 for non-universal perturbations. Universal perturbations do not return average Adversarial Robustness or Attack strength scores, because the trained perturbations still have to be evaluated on a full dataset via evaluate_PCFA.py (see next section).

To specify the datasets and handle how much output is produced, the following arguments help:

--dataset=[Kitti15,Sintel]            | Which dataset is used
--dataset_stage=[evaluation,training] | Use evaluation or training split
--dstype=[clean,final]                | Sintel only, specify clean or final data

--save_frequency=1                    | After so many batches are sample outputs (flow fields, targets, perturbations) written
--no_save                             | Setting this option prevents any but the most important output

Evaluating Existing Perturbations from PCFA

To evaluate perturbations, a path to the perturbation(folder) and the network for which the perturbations were trained have to be specified with the --perturbation_sourcefolder and --origin_net arguments. The --perturbation_sourcefolder can either be the output folder from a universal perturbation training, or a path to a .npy perturbation file. Additionally, the perturbation specification via --joint_perturbation and --universal_perturbation should match the trained patch.

python3 evaluate_PCFA.py --net=[SpyNet,PWCNet,RAFT,GMA,FlowNet2] --origin_net=<specify_network> --perturbation_sourcefolder=<specify> [--joint_perturbation --universal_perturbation]

Further, the dataset on which to evaluate should be specified. Note that perturbations trained for one dataset can only be tested on this dataset (no mixing between KITTI and Sintel perturbations is supported).

Also, the data will be batched according to

--batch_size=4                        | The testing batch size

To specify the datasets and handle how much output is produced, the following arguments help:

--dataset=[Kitti15,Sintel]            | Which dataset is used
--dataset_stage=[evaluation,training] | Use evaluation or training split
--dstype=[clean,final]                | Sintel only, specify clean or final data

--save_frequency=1                    | After so many batches are sample outputs (flow fields, targets, perturbations) written
--no_save                             | Setting this option prevents any but the most important output

Evaluating the Provided Universal Perturbations (Paper Reproducibility)

To reproduce the black-box results (Tab. 4, Supp. Mat. Tab. A4), the best trained universal perturbations are provided for each network. They are in the folder universal_perturbations. To evaluate the effectiveness of a universal perturbation on a new network, specify the universal perturbation as .npy file via the --perturbation_sourcefolder argument, and make sure that --origin_net and --dataset match the universal perturbation.

E.g., to test FlowNet2 with the universal perturbation trained for SpyNet on Kitti15, which is located at universal_perturbations/SpyNet_Kitti15.npy, make sure to specify --origin_net=SpyNet and --dataset=Kitti15, and execute:

python3 evaluate_PCFA.py --net=FlowNet2 --dataset=Kitti15 --dataset_stage=evaluation --origin_net=SpyNet --perturbation_sourcefolder=universal_perturbations/SpyNet_Kitti15.npy --joint_perturbation --universal_perturbation

The universal perturbations for Sintel are generated for Sintel final (--dataset=Sintel --dstype=final).

Training and Evaluating (I-)FGSM perturbations

Additionally, we provide a training routine that can generate FGSM perturbations (image specific only).

python3 attack_FGSM.py --net=[SpyNet,PWCNet,RAFT,GMA,FlowNet2]

Parameters to provide are the number of iterations via --steps, the step size --epsilon as well as the normal attack parameters target and loss.

--steps=20                      | The number I-FGSM steps per image.
--epsilon=0.00025               | The I-FGSM step size

--joint_perturbation            | If this argument is passed, a joint perturbation is trained.

--target=[zero,neg_flow,custom] | Allows to set a target. If a custom target is desired, specify the path to a flow-containing file via --custom_target_path.
--loss=[aee,mse,cosim]          | Allows to specify the loss function

The dataset parameters work as explained for PCFA in the previous sections.

Data Logging and Progress Tracking

Training progress and output images are tracked with MLFlow in mlruns/, and output images and flows are additionally saved in experiment_data/. In experiment_data/, the folder structure is <networkname>_<attacktype>_<perturbationtype>/, where each subfolder contains different runs of the same network with a specific perturbation type.

To view the mlflow data locally, navitage to the root folder of this repository, execute

mlflow ui

and follow the link that is displayed. This leads to the web interface of mlflow.

If the data is on a remote host, the below procedure will get the mlflow data displayed.

Progress tracking with MLFlow (remote server)

Identify the remote's public IP adress via

curl ifconfig.me

then start mlflow on remote machine:

mlflow server --host 0.0.0.0

On your local PC, replace 0.0.0.0 with the public IP and visit the following address in a web-browser:

http://0.0.0.0:5000

Adding External Models

The framework is built such that custom (PyTorch) models can be included. To add an own model, perform the following steps:

Create a directory models/your_model containing all the required files for the model.

Make sure that all import calls are updated to the correct folder. I.e change:

from your_utils import your_functions # old

# should be changed to:
from models.your_model.your_utils import your_functions # new

In helper_functions/ownutilities.py modify the following functions:

import_and_load(): Add the following lines:

 elif net == 'your_model':
 	# mandatory: import your model i.e:
 	from models.your_model import your_model

 	# optional: you can outsource the configuration of your model e.g. as a .json file in models/_config/
 	with open("models/_config/your_model_config.json") as file:
 		config = json.load(file)
 	# mandatory: initialize model with your_model and load pretrained weights
 	model = your_model(config)
 	weights = torch.load(path_weights, map_location=device)
 	model = load_state_dict(weights)

preprocess_img(): Make sure that the input is adapted to the forward pass of your model. The dataloader provides rgb images with range [0, 255]. The image dimensions differ with the dataset. You can use the padder class make the spatial dimensions divisible by a certain divisor.

 elif network == 'your_model':
 	# example: normalize rgb range to [0,1]
 	images = [(img / 255.) for img in images]
 	# example: initialize padder to make spatial dimension divisible by 64
 	padder = InputPadder(images[0].shape, divisor=64)
 	# example: apply padding
 	output = padder.pad(*images)

model_takes_unit_input(): Add your model to the respective list, if it expects input images in [0,1] rather than [0,255].
compute_flow(): Has to return a tensor flow originating from the forward pass of your model with the input images x1 and x2. If your model needs further preprocessing like concatenation perform it here:
```
 elif network == 'your_model':
 	# optional: 
 	model_input = torch.cat((x1, x2), dim=0)
 	# mandatory: perform forward pass
 	flow = model(model_input)
```
postprocess_flow(): Rescale the spatial dimension of the output flow, such that they coincide with the original image dimensions. If you used the padder class during preprocessing it will be automatically reused here.

Add your model to the possible choices for --net in helper_functions/parsing_file.py (i.e. [... | your_model])

External Models and Dependencies

Models

The models provided under models/ are supposed to help recreate the paper results, but are not part of the published attack.

Additional code

Augmentation and dataset handling (datasets.py frame_utils.py InputPadder) from RAFT
Path configuration (conifg_specs.py) inspired by this post
File parsing (parsing_file.py): idea from this post

cv-stuttgart / PCFA