Abstract
Modern recommendation systems rely on sensitive user data, posing significant privacy risks. This paper presents a lightweight, differentially private framework for Matrix Factorization (MF) that protects individual user ratings and profiles from untrusted recommendation engines. By employing objective perturbation and a decoupled computation model supported by a semi-honest third party, the proposed system achieves high accuracy with provable privacy guarantees, even in dynamic environments where users frequently join or leave.
Introduction
Recommendation systems are essential for navigating modern digital marketplaces, yet they typically require users to share private preferences with a central server. Standard Matrix Factorization (MF) techniques—the cornerstone of systems like Netflix's—decompose a user-item rating matrix into latent factor matrices $U$ (user profiles) and $V$ (item profiles). However, access to these matrices allows an untrusted recommender to infer sensitive personal information such as health conditions or political inclinations. While cryptographic methods like garbled circuits offer strong protection, their high computational cost often renders them impractical for real-world use. This research introduces a computationally efficient alternative using Differential Privacy (DP) to secure user data without compromising system utility.
Methodology
The proposed framework utilizes several key techniques to ensure privacy while maintaining performance:
- Objective Perturbation: Unlike standard DP methods that add noise to the final output, this approach injects noise directly into the objective function. This ensures that the learned item profiles ($V$) satisfy differential privacy while maintaining the "direction" of optimization, preserving accuracy.
- Decoupled Computation: User profile ($U$) updates are performed locally on the user's device. The central recommender only receives aggregated, noisy updates, keeping individual profiles secret.
- Noise Decomposition: To ensure the sum of distributed local noise follows the required DP distribution, the authors employ a mathematical decomposition involving Laplace, Exponential, and Normal distributions, allowing users to independently generate "pieces" of noise.
- Mitigating Difference Attacks: To prevent a recommender from canceling out noise by comparing gradients across iterations, a semi-honest third party is introduced. Users encrypt local gradients using a large integer modulo scheme ($mod\ P$). The third party aggregates these masked gradients, so the recommender only sees the final noisy sum.
- Dynamic Support: An initialization phase allows the system to maintain a consistent privacy budget as the set of online users changes over time.
Experimental Evaluation
The framework was evaluated using two industry-standard benchmarks:
- Netflix Dataset: A subset consisting of 191,668 ratings across 100 movies.
- MovieLens 100k: A dataset containing 100,000 ratings from 943 users for 1,682 movies. Both datasets utilize a 5-star rating scale.
Results and Discussion
The experimental results demonstrate that the proposed system maintains high recommendation accuracy while providing robust privacy.
On the MovieLens dataset, the prediction error rapidly approaches the non-private baseline, often within 40 iterations. For the Netflix dataset, the increase in prediction error was minimal—less than 0.03 when the privacy budget $\epsilon \geq 0.15$.
The system also proves efficient for mobile deployment. Communication overhead is low, with users downloading approximately 12KB and uploading 5KB per iteration—well within the capabilities of modern mobile networks. In dynamic settings, the system remains robust, though it requires more iterations (up to 10,000) to converge due to the smaller active user base at any given time.
Conclusion
This research provides a practical path forward for privacy-preserving recommendations. By prioritizing lightweight Differential Privacy over heavy cryptography, the authors demonstrate a system capable of real-world deployment on standard mobile devices and web servers. Future work includes extending this framework to more complex models like Weighted Non-negative Matrix Factorization (WNMF) to further enhance accuracy on sparse datasets.
Summary updated