Privacy-Preserving Neural Networks (PPNN)

A capstone project at Ashoka University (December 2021), completed by Kuber Shahi under the supervision of Dr. Mahavir Jhawar and Dr. Debayan Gupta.

This repository hosts the final report, presentation, proof-of-concept code, and documentation from that work on privacy-preserving machine learning—including a from-scratch neural network trained on MNIST and cryptographic building blocks from SecureNN and SecureML for neural network training in a three-party secure computation (3PC) setting.

---

Capstone deliverables

The main outputs of this project are the written report and presentation. The code in this repo supports and demonstrates the work described in those documents.

Resource	Description
📄 Capstone Report	Full 32-page write-up: ML preliminaries, SecureNN protocol analysis, and the privacy-preserving NN (PPNN) architecture
📊 Capstone Presentation	Slide deck summarizing motivation, methods, and results

Supporting materials

• Research notes — derivations for ReLU, Softmax, Private Compare, DReLU, Division, and related protocols
• Keras reference notebook — same architecture implemented in Python for validation

---

Overview

Modern ML systems need large datasets, but those datasets are often too sensitive to share directly. This project studies how neural networks can be trained using secure multi-party computation (MPC) so that parties can collaborate without revealing raw data.

The work has two complementary tracks:

1. Plain neural network — A C++ baseline (ReLU + Softmax, backprop on MNIST) to understand the math and derive update rules.
2. SecureNN building blocks — Implementations of the MPC primitives needed to run the same network in a privacy-preserving way.

---

Implementation

This repository contains the C++ code referenced in the capstone report.

Plain neural network (build/nn)

• 2-layer fully connected network: 784 → 256 → 10
• ReLU (hidden) and Softmax (output)
• Cross-entropy loss with mini-batch SGD on MNIST (60k train / 10k test)

SecureNN building blocks (build/bb)

Interactive CLI demos of individual protocols:

• Fixed-point mapping and reverse mapping (SecureML-style)
• Truncation after ring arithmetic
• Additive secret sharing in ℤ_L and ℤ_p
• Secure matrix multiplication via Beaver triples
• Private Compare (unshared and shared settings)

---

Project structure

ppnn-capstone/
├── docs/
│   ├── Capstone_Report.pdf          # primary deliverable
│   ├── Capstone_Presentation.pdf    # primary deliverable
│   ├── assets/                      # diagrams for README
│   └── notes/                       # protocol derivations
├── src/                             # C++ implementation
├── archive/                         # early development scratch (not built)
├── notebooks/                       # Python/Keras reference
├── datasets/mnist.zip
├── scripts/setup-dataset.sh
├── Makefile
└── README.md

---

Quick start

git clone https://github.com/kubershahi/ppnn-capstone.git
cd ppnn-capstone

chmod +x scripts/setup-dataset.sh
./scripts/setup-dataset.sh

make
./build/nn    # train plain NN on MNIST
./build/bb    # demo SecureNN building blocks

Requirements: C++11 compiler, Eigen3, make, unzip

# macOS
brew install eigen

# Ubuntu / Debian
sudo apt-get install libeigen3-dev

If the build can't find Eigen, pass the include path explicitly:

make EIGEN_INCLUDE=-I/usr/include/eigen3

---

References

• Mohassel & Zhang, SecureML
• Wagh et al., SecureNN
• Goodfellow, Bengio & Courville, Deep Learning

---

License

This project is licensed under the MIT License.

The capstone report and presentation are academic work from Ashoka University (2021). If you use or build on this project, please cite the report and credit the original authors.