SAR-ATR-TransAttack

This repository provides the official implementation for the paper:

"SRAW-Attack: Space-Reweighted Adversarial Warping Attack for SAR Target Recognition"

The codebase is designed to support reproducible research on adversarial robustness and vulnerability analysis of deep learning–based SAR Automatic Target Recognition (SAR-ATR) systems, with a particular focus on spatial/warping-based adversarial attacks that depart from conventional additive perturbation paradigms.

---

1. Overview

Deep neural networks (DNNs) have become the dominant approach for SAR Automatic Target Recognition (ATR). However, due to the intrinsic sparsity, strong structural priors, and speckle-dominated background of SAR imagery, their robustness characteristics differ fundamentally from those observed in optical imagery.

This repository investigates warping-based adversarial attacks for SAR-ATR, including spatial warping and geometry-aware perturbations, and provides:

• Implementations of several baseline adversarial attacks adapted to SAR imagery
• A novel warping-based adversarial attack framework (TransAttack)
• Evaluation protocols tailored for SAR-ATR models
• Experimental code supporting the results reported in the paper

The overall goal is to analyze, rather than merely degrade, SAR-ATR robustness, and to better understand how geometric distortions and pixel re-arrangements can serve as effective adversarial mechanisms in SAR.

---

2. Key Contributions Reflected in This Code

• Beyond additive perturbations
  Unlike FGSM/PGD-style methods that rely on pixel-wise intensity changes, this work formulates adversarial generation as a spatial transformation problem, which is particularly suitable for SAR images with strong structural cues.

• SAR-specific adversarial insights
  The implementation explicitly considers SAR characteristics such as sparse scattering centers, background speckle, and geometry sensitivity, making it fundamentally different from off-the-shelf vision attack code.

• Unified evaluation pipeline
  The repository supports consistent comparison between classical additive attacks and transformation-based attacks under the same SAR-ATR setting.

---

3. Repository Structure

SAR-ATR-TransAttack/
├── attacks/            # Adversarial attack implementations
│   ├── fgsm.py
│   ├── pgd.py
│   ├── cw.py
│   ├── ...
│   └── sraw.py  # Transformation-based attack (core contribution)
│
├── models/             # SAR-ATR network architectures
│   ├── resnet.py
│   ├── vgg.py
│   └── ...
│
├── datasets/           # Dataset loading and preprocessing
│   └── mstar.py
│
├── utils/              # Helper functions (metrics, visualization, etc.)
│
├── configs/            # Experiment configuration files
│
├── train.py            # Model training script
├── test.py             # Standard evaluation
├── attack_eval.py      # Adversarial evaluation entry point
│
└── README.md

Note: The exact structure may evolve as experiments are extended. The current layout reflects the version used for the paper's reported results.

---

4. Supported Attacks

The following adversarial methods are implemented or supported:

• FGSM (Fast Gradient Sign Method)
• PGD (Projected Gradient Descent)
• MI-FGSM
• CW Attack
• SRAW Attack (proposed)

The proposed SRAW optimizes adversarial effectiveness by learning spatial deformation fields rather than direct pixel perturbations, enabling subtle but semantically meaningful attacks in SAR imagery.

---

5. Datasets

Experiments are conducted on SAR-ATR benchmark datasets (e.g., MSTAR-style datasets).

Due to dataset licensing restrictions, raw SAR data are not included in this repository. Please prepare the dataset separately and organize it according to the expected directory structure defined in datasets/.

---

6. Usage

6.1 Environment Setup

conda create -n sar-attack python=3.8
conda activate sar-attack
pip install -r requirements.txt

6.2 Training a SAR-ATR Model

python train.py --config configs/train.yaml

6.3 Adversarial Evaluation

python attack_evaluation.py \
    --attack transattack \
    --model resnet \
    --config configs/attack.yaml

---

7. Evaluation Metrics

The code supports commonly used metrics for adversarial evaluation in SAR-ATR, including:

• Classification accuracy under attack
• PSNR / SSIM
• Perceptual similarity metrics
• Attack success rate (ASR)

These metrics are reported in the paper to quantify both attack effectiveness and visual plausibility.

---

8. Reproducibility

• All experiments are conducted with fixed random seeds.
• Configuration files used for the paper are provided in configs/.
• Model checkpoints can be regenerated following the provided scripts.

If you encounter discrepancies, please verify dataset preprocessing and normalization settings, which are critical for SAR data.

---

9. Citation

If you find this work useful in your research, please consider citing:

@article{SRAW,
  title={SRAW-Attack: Space-Reweighted Adversarial Warping Attack for SAR Target Recognition},
  author={Yiming Zhang and others},
  journal={Arxiv},
  year={2026}
}

---

10. License

This project is released for academic research purposes only.

Please check the license file for details and ensure proper citation when using this code in your work.

---

11. Contact

For questions, discussions, or potential collaborations, feel free to open an issue or contact the authors.

---

This repository aims to facilitate transparent, reproducible, and SAR-aware adversarial research.