Privacy-Preserving Machine Learning (PPML)

Research implementation and evaluation of secure multi-party machine learning protocols, developed during a research internship at Mphasis Lab, Ashoka University (May–August 2021).

Advisor: Prof. Mahavir Jhawar

Overview

This repository contains implementations and experiments for Privacy-Preserving Machine Learning (PPML) — training and evaluating ML models on sensitive data without exposing raw inputs to any single party.

The work focuses on:

1. SecureML — A C++ implementation of the two-party secure training protocol from Mohassel & Zhang (IEEE S&P 2017), supporting linear and logistic regression on tabular and image datasets.
2. BLAZE — A Python prototype of the three-party BLAZE framework (Kumar et al., 2020) with secure multiplication, truncation, and sharing primitives.
3. Novel protocol research — Exploratory work on optimized activation functions (Double ReLU) and fixed-point arithmetic for business-specific PPML use cases.

The goal was to compare protocol efficiency and security trade-offs, then extend the most promising approaches for real-world deployment scenarios.

---

Repository Structure

mphasis_ppml/
├── assets/            # README figures and diagrams
├── secureml/          # C++ SecureML implementation (linear & logistic regression)
├── blaze/             # Python BLAZE protocol primitives and tests
├── novel/             # Custom protocol research (Double ReLU, fixed-point MPC)
├── datasets/          # Sample datasets and preprocessing notebooks
├── experiments/       # Auxiliary prototypes and unit tests
├── bin/               # Compiled executables (generated)
└── build/             # Object files (generated)

Directory	Description
assets/	Overview diagram and other documentation figures
secureml/	Two-party secure ML training in C++ with Eigen
blaze/	Three-party BLAZE protocol in Python
novel/	Research notes and prototypes for custom PPML extensions
datasets/	Medical insurance splits, MNIST archive, preprocessing notebooks
experiments/	Truncation tests, OT experiments, and helper scripts

---

Getting Started

Prerequisites

SecureML (C++)

• Eigen — linear algebra
• g++ with C++14 support

If Eigen is not on the default include path (common on macOS with Homebrew):

make EIGEN_INCLUDE=/opt/homebrew/include/eigen3

BLAZE (Python)

• Python 3.7+
• numpy, gmpy2 — install via pip install -r blaze/requirements.txt

Experiments (optional)

• emp-ot — only required for experiments/ot-test.cpp

Build & Run SecureML

From the repository root:

make linear    # secure linear regression
make logistic  # secure logistic regression

./bin/linear    # run from repository root
./bin/logistic

Both programs prompt for a dataset:

Option	Dataset
1	MNIST (requires extraction — see datasets/README.md)
2	Medical insurance (included)
3	Binary MNIST (requires generation — see Binary MNIST generation.ipynb)
4	Sanity-check toy data

Run BLAZE Tests

pip install -r blaze/requirements.txt
cd blaze
python3 main.py    # primitive and sharing semantics tests
python3 main2.py   # extended protocol tests

---

Protocols & References

Protocol	Setting	Paper
SecureML	2-party, semi-honest	Mohassel & Zhang, IEEE S&P 2017
BLAZE	3-party, honest majority	Kumar et al., 2020

Key Techniques

• Secret sharing — Additive and special (three-party) sharing schemes
• Beaver triples — Secure multiplication without revealing inputs
• Fixed-point arithmetic — Integer-ring embeddings with configurable precision (13–18 bit scaling)
• Truncation — Post-multiplication fixed-point reduction to control overflow

---

Limitations

This is research prototype code, not a production-ready secure ML framework:

• Semi-honest security assumptions; no formal side-channel hardening
• Fixed-point precision and truncation parameters are dataset-dependent
• MNIST and binary MNIST require manual setup (see datasets/)
• Some BLAZE networking paths are stubbed for local testing

---

License

Released under the MIT License.