This repository provides a minimal, reproducible evaluation package for the corresponding research paper.
It includes pretrained checkpoints, evaluation scripts, and robustness/attack evaluation code so that reviewers and readers can reproduce the main results without retraining models.
replicate-package/
├── checkpoints/
│ ├── ckpt_recon/
│ │ └── 30/
│ │ ├── epsl-1_c.pt
│ │ └── epsl-1_g.pt
│ │
│ └── ckpt_attack_recon/
│ └── 30/
│ └── epsl-mixed_g.pt
│
├── evaluation/
│ ├── utils.py
│ ├── reconstruction.py
│ ├── attack_reconstruction.py
│ ├── attack_fgsm.py
│ └── robust_gaussian.py
│
├── models/
│ ├── sensing.py
│ └── generator.py
│
├── scripts/
│ ├── eval_reconstruction.py
│ ├── eval_attack_recon.py
│ ├── eval_attack_fgsm.py
│ └── eval_robust_gaussian.py
│
├── requirements.txt
└── README.md
conda create -n privcs python=3.9 -y
conda activate privcs
pip install -r requirements.txtAlways run scripts from the repository root using module mode:
python -m scripts.eval_reconstructionpython -m scripts.eval_reconstruction \
--dataset cifar10 \
--ckpt_root checkpoints/ckpt_recon \
--compression_rates 30 \
--epsilons 1 \
--batch_size 256python -m scripts.eval_attack_recon \
--dataset cifar10 \
--ckpt_root checkpoints \
--compression_rates 30 \
--batch_size 256python -m scripts.eval_attack_fgsm \
--dataset cifar10 \
--ckpt_root checkpoints/ckpt_recon \
--compression_rates 30 \
--epsilon 1 \
--adv_scale 0.1python -m scripts.eval_robust_gaussian \
--dataset cifar10 \
--ckpt_root checkpoints/ckpt_recon \
--compression_rates 30 \
--epsilon 1 \
--gaussian_scale 0.20- PSNR / SSIM
- SSIM uses win_size = 3 for CIFAR‑10
Only selected checkpoints are uploaded to ensure reproducibility while keeping the repository lightweight.
Please open an issue if you encounter any problem during reproduction.