Connecting the Dots in Privacy-Preserving ML

This repository contains supplementary material for the paper
“SoK: Connecting the Dots in Privacy-Preserving ML — Systematization of MPC Protocols and Conversions Between Secret Sharing Schemes.”

---

Full Version

The full version of our paper is available as a preprint on the Cryptology ePrint Archive: https://eprint.iacr.org/2025/1679.

---

Contents

📚 Comprehensive Survey of Related Work

We highlight the main differences between existing surveys and SoKs in the domain of privacy-preserving machine learning (PPML).

• Detailed discussion is provided in Appendix A of the full paper.
• A consolidated overview can be found here.

---

⚙️ Design & Deployment Dimensions

We systematize MPC-based PPML protocols along key dimensions:

• Algebraic structure
• Threat model
• Execution phase
• Deployment mode
• Network

This analysis highlights the trade-offs between efficiency and security.

• Detailed discussion is provided in Appendix C of the full paper.
• A comprehensive table classifying considered frameworks across all dimensions can be viewed here.
• We further provide high-level categorization based on the MPC techniques used, support for ML training or inference, or availability of either theoretical or experimental evaluation. We split the tables based on the number of parties: 2PC, 3/4PC, nPC

---

🤖 ML-Based Systematization and 🔐 Low-Level Protocol Analysis

We categorize frameworks based on their support for different ML functionalities in Neural Networks and Transformer models.

• Detailed discussion is provided in Appendix D of the full paper.
• The overview of supported functionalities is split based on the number of parties: 2PC and MPC

We further decompose PPML frameworks into their core cryptographic primitives and provide a comprehensive overview of the theoretical costs for different ML functionalities. We focus on the most common functionalities, with concrete costs and approaches detailed in corresponding tables:

• Dot-Product
• Truncation
• ReLU
• Softmax
• Sigmoid
• GELU
• Normalization

---

🧩 Unification and Conversions

Through the MPC Puzzle, we unify 2-, 3-, and 4-party secret-sharing schemes and present conversion protocols among them, including an analysis of their communication costs.

• Detailed discussion is provided in Appendix E of the full paper.

---

Citation

Please cite as:

@article{ZbudilaSYMAP25,
  author       = {Martin Zbudila and
                  Ajith Suresh and
                  Hossein Yalame and
                  Omid Mirzamohammadi and
                  Aysajan Abidin and
                  Bart Preneel},
  title        = {{SoK: Connecting the Dots in Privacy-Preserving {ML} - Systematization
                  of {MPC} Protocols and Conversions Between Secret Sharing Schemes}},
  journal      = {{IACR} Cryptol. ePrint Arch.},
  year         = {2025},
  url          = {https://eprint.iacr.org/2025/1679}
}