Code and videos accompanying the paper "Over-the-Air Adversarial Flickering Attacks against Video Recognition Networks".

We encourage the readers to view the adversarial videos in the following:

Our threat models follows the white-box setting. In the experiments we used the I3D , R(2+1)D, R3D and MC3 as target models (pre trained on the Kinetics dataset ).

This repository contains the code to reproduce our reported results.

This code has been tested on Ubuntu 16.04, Python 3.7.5, Tensorflow 1.15, torch 1.4.0 and Titan-X GPUs.

• Clone the repository

git clone https://github.com/anonymous-p/Flickering_Adversarial_Video
cd Flickering_Adversarial_Video

• Install the dependencies

pip install -r requirements.txt

• Download and merge checkpoint and additional data

./download_ckpt_and_data.sh
mv data/result/ .

We provide 3 runnig scripts:

1. Single Video Attack
2. Single Class Generalization Attack
3. Universal Attack

The attack's running configuration can be modified by the file run_config.yml

Configuration section SINGLE_VIDEO_ATTACK in run_config.yml

• Run adversarial_main_single_video_npy.py

Flickering_Adversarial_Video$ python i3d_adversarial_main_single_video_npy.py

• Visualize the adversarial video and perturbation

The result file (.pkl) will be save according to PKL_RESULT_PATH field in run_config.yml.

We provide example of .pkl:

Flickering_Adversarial_Video$ python utils/stats_and_plot/stats_plots.py result/videos_for_tests/npy/bartending_beta1_0.1_th_1.67%_rg_1.19%.pkl

• Download kinetics databse according to Kinetics-Downloader
• Convert the downloaded data to .tfrecord:

 python kinetics_to_tf_record_uint8.py <data_dir> class <tfrecord_data_dir> 

<data_dir> : the source (.mp4) data dir

class : class to convert

<tfrecord_data_dir> : destination path to save .tfrecord's

for example, convert only triple jump in val split to tfrecord:

 python kinetics_to_tf_record_uint8.py 'data/kinetics/database/val/' 'triple jump' 'data/kinetics/database/tfrecord_uint8/val/' 

for example, convert all class in val split to tfrecord:

 python kinetics_to_tf_record_uint8.py 'data/kinetics/database/val/' 'all' 'data/kinetics/database/tfrecord_uint8/val/' 

Configuration section CLASS_GEN_ATTACK in run_config.yml

• Run adversarial_main_single_class_gen.py

Flickering_Adversarial_Video$ python i3d_adversarial_main_single_class_gen.py

• Download kinetics databse according to Kinetics-Downloader
• Convert the downloaded data to .tfrecord s.t create tfrecord's with shuffled classes':

 python kinetics_to_tf_record_uint8_shuffle.py 'data/kinetics/database/test/' 'data/kinetics/database/tfrecord/test_all_cls/' 

Configuration section UNIVERSAL_ATTACK in run_config.yml

• Run adversarial_main_universal.py

Flickering_Adversarial_Video$ python i3d_adversarial_main_universal.py

In order to run Sparse Adversarial Perturbations for Videos set FLICKERING_ATTACK to False

Loss's and metric's can be monitor with 'tensorboard', for example:

Flickering_Adversarial_Video$ tensorboard --logdir=result/generalization/universal_untargeted/

We provide 2 runnig scripts:

1. Single Video Attack
2. Universal Attack

Before using the running scripts the kinetics databse shuold be downloaded according to Kinetics-Downloader.

The attack's running configuration can be modified inside the running script under 'MODEL AND TRAIN PARAMETERS', allowing to set the attacked model ( R(2+1)D, R3D and MC3) and other parameters regarding the attack.

• Run r2plus1d_main_statistics_single_video_attack.py

Flickering_Adversarial_Video$ python r2plus1d_main_statistics_single_video_attack.py

• Run r2plus1d_main_universal_attack.py

Flickering_Adversarial_Video$ python r2plus1d_main_universal_attack.py