SEGA: Signed Ensemble Gaussian Black-box Attack against NR-IQA Models

This repository contains the code and resources for reproducing the main experiments in the TPAMI paper "SEGA: A Transferable Signed Ensemble Gaussian Black-box Attack Method for No-Reference Image Quality Assessment Models."

File Overview

• SEGA.py: Main implementation of the SEGA attack.
• test_transferability.py: This script evaluates attack performance across different target models. Note: Results are generated directly from float image data and may show minor variations from the paper's reported results, which are based on saved .bmp images.
• threatmodels.py: Definitions of NR-IQA threat models used in the experiments.
• threatmodels_class.py: Model classes for NR-IQA models.
• livec-test.csv: A CSV file containing image ids and quality labels for testing on the CLIVE dataset.
• checkpoints/: Directory containing pretrained model checkpoints required for running the attack.

Dependencies

• Python >= 3.7
• PyTorch >= 1.8
• NumPy
• OpenCV
• scikit-image

Before Running

• Download the LIVEC dataset and place the center-cropped images in the following directory:
  'YOUR/PATH/TO/CLIVE/CENTERCROPPED/IMAGES'
  Then, update line 89 in test_transferability.py with this path.
• Download the pretrained checkpoints from the provided google drive URL or baidu pan URL and place them in the checkpoints/ directory.

How to Use

To run the transferability evaluation, use the following command:

python test_transferability.py --target_model DBCNN --source_model_set LIQE,HyperIQA,LinearityIQA --num_samples 10 --smooth_std 10./255