lixi1994/temporal-distributed-backdoor-attack-against-video-based-action-recognition

TDBA-VAR

Codes for the paper "Temporal-Distributed Backdoor Attack Against Video Based Action Recognition".

Requirements

pip install -r requirements.txt

Dataset

We used the UCF101 dataset for our project. Put the data to 'UCF-101-imgs'.

Run

cd code

Train slowfast model poisoned by DFT based attack

python train_adv.py --attack FFT_slowfast --model_type slowfast

Train s3d/i3d/res2+1d model poisoned by DFT based attack

python train_adv.py --attack FFT_downsample --model_type s3d

temporal-distributed-backdoor-attack-against-video-based-action-recognition's Issues

i3d model problem

There are some problems with I3D model below.

1. Model avgpool(x) in line 439 (/src/i3d/i3d.py)
   RuntimeError: input image (T: 4 H: 8 W: 10) smaller than kernel size (kT: 8 kH: 7 kW: 7)

I guess last_duration = int(math.ceil(num_in_frames / 8)) is 8 (kT), and your input is 4.

2. If I revise the last_duration as 2. The new error message is
   RuntimeError: only batches of spatial targets supported (3D tensors) but got targets of size: : [16]

I guess the size of output from the model is torch.Size([16, 1024, 4, 8, 10]), but targets of size: : [16] incompatible.

So, I3D in your code can not successfully train. If you fix this problem, I will appreciate that.

Could you provide the prerequisite for reproduction?

I'd like to extend my sincere appreciation for your outstanding contributions to this work. I am keenly interested in your work and have been endeavoring to reproduce it in my project, but have encountered some difficulties.

Could you provide the necessary guidance or share the required environment settings and dependencies? For instance, please share the package (requirement.txt) and dataset download links used in your work. In addition, if you could provide the folder structure, it would greatly assist me in my endeavors.

Furthermore, if you have any insights or tips on how to effectively tackle this issue, I would be immensely grateful.

Thank you once again for your remarkable efforts.