Unrevealed Threats: Adversarial Robustness Analysis of Underwater Image Enhancement Models

[CODE] [PAPER]

Environment setup

pip install -r requirements.txt

Data preparation

Download underwater image datasets from UIEB and EUVP. Follow the data organization from below and config in dataset_path_config.py

├── dataset_name
    ├── train
        ├── images
            ├── im1.jpg
            ├── im2.jpg
            └── ...
        ├── labels
            ├── im1.jpg
            ├── im2.jpg
            └── ...
    ├── val
        ├── images
            ├── im1.jpg
            ├── im2.jpg
            └── ...
        ├── labels
            ├── im1.jpg
            ├── im2.jpg
            └── ...

Training undefended models.

cd adv_train
python train_ADMNNet.py --dataset_name UIEB --name not_adv --results_dir ../results_models/UWIE/  --train_batch_size 6

Adversarial attacks

To conduct adversarial attacks, use the following instructions.

cd adv_eval
python adv_ADMNNet.py --dataset_name UIEB --model_path ../results_models/UWIE/ADMNNetnot_adv_100/UIEB/models/last_ADMNNet_UIEB.pth --results_dir ../results_UWIE_adv_eval/UWIE/ --train_batch_size 6

Adverarial training

To evaluate the robustness of defended models, use the following instructions.

# finetune
cd adv_train
python train_ADMNNet.py --finetune \
                        --adv_train --name adv_f \
                        --num_epochs 20 \
                        --dataset_name UIEB --results_dir ../results_models/UWIE \
                        --train_batch_size 6

# training from scratch
python train_ADMNNet.py  
--adv_train --name adv_s \
--num_epochs 100 \
--dataset_name UIEB --results_dir ../results_models/UWIE \
--train_batch_size 6