ThesisResearchSecurity 10-14 months
Privacy-Preserving Recommender System using Homomorphic Encryption
A recommendation system where the server computes personalized recommendations on encrypted user preference data — using homomorphic encryption so user privacy is mathematically guaranteed even if the server is compromised.
PythonTenSEAL (CKKS scheme)PyTorchFastAPINumPy
Find GitHub repos & source code →How to build it — step by step
- 1HE Basics: Study CKKS scheme for approximate arithmetic on encrypted vectors; implement basic add/multiply on ciphertexts
- 2Matrix Factorization: Implement encrypted SVD/ALS for collaborative filtering; all dot products computed in ciphertext domain
- 3Protocol Design: Client encrypts preference vector; sends to server; server computes recommendations on ciphertext; returns encrypted result
- 4Performance: Measure recommendation quality (NDCG, MAP) vs plaintext baseline; measure encryption overhead
Key features to implement
- ✓Mathematically proven privacy — server cannot see user data
- ✓Accurate recommendations despite encryption
- ✓Client-side key management
- ✓Performance analysis vs plaintext systems
- ✓Practical deployment guide for HE recommenders
💡 Unique twist to stand out
Compare CKKS homomorphic encryption with Secure Multi-Party Computation (SMPC) approaches for the same recommendation task — providing a practical guide on when to prefer HE vs SMPC for privacy-preserving ML.
🎓 What you'll learn
Homomorphic encryption theory and practice, privacy-preserving machine learning, recommender system design, and cryptographic protocol engineering.