AO-Exp-Attack

Framework

Video-based object detection is essential in safety-critical systems such as autonomous driving and surveillance. While modern deep learning-based object detectors have achieved remarkable accuracy, they remain vulnerable to adversarial attacks — particularly universal perturbations that generalize across multiple inputs.

This project presents a novel, minimally distorted universal adversarial attack designed specifically for video object detection.

Universal adversarial perturbations (UAP)

Getting Started

Tested on Linux Ubuntu Ubuntu 22.04.5 LTS with python 3.10.12.

First clone repository:

git clone https://github.com/jsve96/AO-Exp-Attack.git

Set up virtual environment:

python -m venv UAT
source ~/path/to/AO-Exp/bin/activate
pip install torch==2.6.0 torchvision==0.21.0 torchaudio==2.6.0 --index-url https://download.pytorch.org/whl/cu124
pip install -r requirements.txt

You can place the dataset wherever you prefer. If you do so, then update:

DATA_PATH = ["path/to/dataset"]

Usage

If you want to run AO-Exp Attack on the VTM dataset then navigate to Repository and run:

python3 methods/AOEXPAttackVTMBATCH.py

Hyperparameter

Parameter	Description	Value / Condition
dataset	Name of the dataset	"VTM-Data"
path_dataset	Path to the dataset (from DATA_PATH, a list or iterable)	e.g. DATA_PATH[0]
instance	Video instance name	"Video1" to "Video15"
nit	Number of iterations	50 if k == 'top1', else 100
l1	Nuclear norm regularization	0.1
l2	Frobenius regularization	0.001 if k == 'top1', else 0.0002
save	Path to save experiment outputs	REPO_PATH/experiments/VTM-Data/<instance>/AO-Exp
k	Attack mode: top-1 targeting or scalar multiplier	'top1' or 1.0
DEVICE	Device used for computation	torch.device('cuda:0')

You can change all hyperparameter in /methods/AOEXPAttackVTMBATCH.py by simply modifying

if __name__ == "__main__":
    from itertools import product
   
    Data = ['VTM-Data']
    instance = ['Video'+ str(i) for i in range(1,16)]
    Path_Data = DATA_PATH
    k = [1.0, 'top1']
    L1 = [0.1]
    DEVICE = torch.device('cuda:0')

    

    for dataset,path_dataset,l1,instance,k in product(Data,Path_Data,L1,instance,k):
        args = argparse.Namespace(
            dataset=dataset,
            path_dataset = path_dataset,
            nit = 50 if k == 'top1' else 100,
            l1=l1,
            l2=0.001 if k =='top1' else 0.0002,
            save=REPO_PATH+'/experiments/VTM-Data/'+instance+"/AO-Exp",
            instance = instance,
            k=k,
            DEVICE = DEVICE
        )
        print(args.dataset)
        main(args) 

If you want to run all experiments you need to run all python scripts in /methods.