Encrypted Collaborative Filtering for Privacy-Preserving Recommendation

• This project is made for final-project of SIT 24F CPE602-Applied Discrete Math.
• This project implements a privacy-preserving collaborative filtering recommendation system using encryption techniques, secret sharing, and zero-knowledge proof validation. The core objective is to ensure the security and privacy of user and item data while maintaining high recommendation accuracy.

Project Overview

• Framework: PyTorch
• Data Source: MovieLens dataset (ml-latest-small)
  • https://grouplens.org/datasets/movielens/
• Core Techniques:
  • AES encryption for privacy-preserving computation
  • Zero-Knowledge Proof (ZKP) for data integrity
  • Enhanced matrix factorization for recommendation

Features

1. Privacy-Preserving Computation:

   • User and item IDs are encrypted using AES encryption.
   • Encrypted data ensures confidentiality during computations.

2. Zero-Knowledge Proof:

   • Validates the integrity of secret-shared data.

3. Enhanced Matrix Factorization:

   • Incorporates user and item embeddings with a deep neural network for accurate predictions.

4. Performance Metrics:

   • Tracks training loss, testing loss, RMSE, and MAE during model training.
   • Visualizes performance metrics over training epochs.

Dependencies

Ensure the following dependencies are installed:

Pandas==1.3.5
Scikit-learn==1.0.2
PyTorch==1.7.1
Matplotlib==3.5.3

Sample Results

Epoch 1: Train Loss=1.3476, Test Loss=0.1851, RMSE=0.4302, MAE=0.3394
...
Epoch 30: Train Loss=0.0432, Test Loss=0.0356, RMSE=0.1886, MAE=0.1471

Sample Predictions:
Actual: [0.6 0.7 0.8 0.8 0.7], Predicted: [0.5351176  0.7222369  0.8366588  0.7901543  0.64153785]

Challenges and Decisions

During development, we considered using frameworks like PySyft, CrypTen, and TF-Encrypted for privacy-preserving computations. However, we encountered the following challenges:

1. Hardware and Environment Requirements:

   • These frameworks require specific hardware configurations and software environments, such as particular versions of Python, PyTorch, or TensorFlow. Incompatibilities with our current setup made their usage impractical.

2. Learning Curve and Performance Limitations:

   • PySyft and CrypTen abstract the complexities of multi-party computation and encryption but involve a steep learning curve and debugging challenges.
   • TF-Encrypted has slower updates and limited efficiency, making it unsuitable for the performance demands of this project.

Why Manual Implementation?

Instead of relying on these frameworks, we manually implemented AES encryption and zero-knowledge proof mechanisms. While not a fully secure encryption system, this approach:

• Simplified hardware and environment dependencies, ensuring ease of execution.
• Provided deeper insights into encryption principles and multi-party computation.
• Enhanced flexibility for experimenting with different configurations and scenarios.

File Structure

• main.py
• Dataset/
• README.md
• LICENSE

License

This project is licensed under the MIT License. See the LICENSE file for details.

---

For any questions or contributions, feel free to reach out!