liushiyi1994 / Video-Perturbation

Code for NDSS paper: Stealthy Adversarial Perturbations Against Real-Time Video Classification Systems

Video-Perturbation

Code for NDSS paper: Stealthy Adversarial Perturbations Against Real-Time Video Classification Systems (https://arxiv.org/pdf/1807.00458.pdf)

Requirements:

1. Have installed the tensorflow >= 1.2 version

2. You must have installed the following two python libs: a) tensorflow b) Pillow c) numpy d) cv2

3. You must have downloaded the UCF101 (Action Recognition Data Set)

4. Each single avi file is decoded with 5FPS in a single directory. (ref: https://github.com/hx173149/C3D-tensorflow)

   • you can use the ./list/convert_video_to_images.sh script to decode the ucf101 video files
   • run ./list/convert_video_to_images.sh .../UCF101 5

Usage:

1. Data prepare

1. python data_prepare.py will process (crop, subtract image by mean-image, ...) the images from UCF 101 dataset and save one video clip as one .npy file for the later convenience for loading data.

2. UCF 101 train test split 1 is used.

2. Model

1). python train.py will train the C-DUP or 2D-DUP generator. The trained model will be saved in ./G_model_3D or ./G_model_2D directory.

2). python generate_pertubation.py will generate perturbations from the C-DUP generator and save the perturbations in ./G_model_3D or generate perturbations from the 2D-DUP generator and save the perturbations in ./G_model_2D directory.

3). python test.py will test the attack success rate with the generated perturbation.

Trained model:

C3D model, C-DUP generator, and 2D-DUP generator are available at Dropbox: https://www.dropbox.com/sh/mmq4922i0llok8d/AABPAuqLwN-m1465kDz3QhG1a?dl=0

Experiment results:

See the paper