Threat-Driven Frameworks for Privacy-Preserving Machine Learning (2017–2025)

Author: Mohammed Saad Shareef, Syed Ahsan Ahmed
Lords Institute of Engineering and Technology, Hyderabad, India

One-Page Research Summary
For a quick overview, reviewers and professors can read the concise 1-page summary:

• Download Research Summary (PDF)
• Download Preprint (PDF)

---

Overview

This repository accompanies the preprint:

Benchmarking the Practical Adoption of Privacy-Preserving Machine Learning:
A Tool-Centric and Framework-Oriented Review (2017–2025)

The repository contains all supplemental material required to reproduce the DP-SGD experiment (Appendix A), generate figures, and reference framework-selection tools such as the practitioner decision tree and threat→defense map.

It is designed for:

• Researchers evaluating PPML tools (Opacus, TF-Privacy, FATE, FLARE, CrypTen, SEAL, Flower)
• Practitioners comparing DP, FL, SMPC, HE, and hybrid frameworks
• Students studying privacy–utility tradeoffs
• Reviewers verifying reproducibility

---

Repository Structure

Path	Description
run_privacy_accounting.py	Canonical DP-SGD MNIST experiment (Appendix A).
plot_epsilon_vs_accuracy.py	Minimal IEEE-style privacy–utility curve generator.
notebooks/privacy_accounting.ipynb	Interactive demo (quick mode) calling the script.
figures/	All final paper figures (threat→defense map, decision tree, Appendix plot).
results/	Processed CSV results used in figures (eps_results_final.csv or demo CSV).
paper/	LaTeX source + preprint version (replace with final PDF).
decision_matrix.csv	Practitioner-oriented comparison of PPML frameworks (DP, FL, HE, SMPC).
requirements.txt	Reproducible environment for all scripts and notebooks.

Raw datasets (MNIST) are not committed and are downloaded automatically via torchvision.

---

Quick Start (Reproduce Appendix A)

1 Install dependencies

python -m venv venv source venv/bin/activate # Windows: venv\Scripts\activate pip install -r requirements.txt

2 Run the DP-SGD experiment (quick demo)

This runs 1 epoch, 1 seed, sigma=1.0 → fast reproducibility check.

python run_privacy_accounting.py --quick --out results/eps_results_demo.csv

3 Generate the Privacy–Utility Curve

python plot_epsilon_vs_accuracy.py

Output saved to: figures/epsilon_vs_accuracy.png figures/epsilon_vs_accuracy.svg

4 Full reproduction (slower; full Appendix A)

python run_privacy_accounting.py --epochs 15 --batch 128 --out results/eps_results_final.csv python plot_epsilon_vs_accuracy.py

Decision Matrix (Practitioner Tool)

decision_matrix.csv compares major PPML frameworks across:

• Threat coverage (MIAs, inversion, extraction, leakage, poisoning)
• Techniques (DP, FL, SMPC, HE, hybrid)
• Maturity + ecosystem support
• Performance + scalability
• Real-world adoption
• Documentation & ease-of-use

Each entry includes a provenance column clarifying whether values come from:

• paper experiments (paper_run)
• literature surveys (literature)
• framework documentation (tool_doc)
• external references (repo_link)

Experimental Reproducibility

This repository includes:

• Deterministic seeds for all runs
• CSV logs for DP-SGD experiments
• Full plotting script (IEEE-style)
• A lightweight demo notebook
• Clear environment + version pinning

Hardware used for primary results:

• GPU: Tesla T4 / RTX series
• PyTorch ≥ 2.0
• Opacus ≥ 1.1

Citation

If you use this repository, please cite:

@article{shareef2025ppmlsurvey, title={Benchmarking the Practical Adoption of Privacy-Preserving ML: A Tool-Centric and Framework-Oriented Review (2017--2025)}, author={Ahmed,Syed Ahsan and Shareef,Mohammed Saad}, year={2025}, journal={Preprint/ Workshop Submission}, }

CITATION.cff is also provided.

License

This project is released under the MIT License. You are free to reuse, modify, and distribute with attribution.

Contributing

Pull requests are welcome for:

• Extending the decision matrix
• Adding new PPML frameworks
• Improving plotting or reproducibility tools
• Reporting errors in the threat→defense mappings

Contact

For collaboration or research inquiries:

• Syed Ahsan Ahmed
• Mohammed Saad Shareef