Survey of Privacy-Preserving Collaborative Filtering

Islam Elnabarawy, Wei Jiang, Donald C. WunscharXiv 2020 · DOI · arXiv

Abstract

Collaborative filtering recommendation systems provide recommendations to users based on their own past preferences, as well as those of other users who share similar interests. The use of recommendation systems has grown widely in recent years, helping people choose which movies to watch, books to read, and items to buy. However, users are often concerned about their privacy when using such systems, and many users are reluctant to provide accurate information to most online services. Privacy-preserving collaborative filtering recommendation systems aim to provide users with accurate recommendations while maintaining certain guarantees about the privacy of their data. This survey examines the recent literature in privacy-preserving collaborative filtering, providing a broad perspective of the field and classifying the key contributions in the literature using two different criteria: the type of vulnerability they address and the type of approach they use to solve it.

Introduction

Collaborative Filtering (CF) is the foundational technology behind modern recommendation engines like Netflix or Amazon. By comparing a user's history with that of similar peers, these systems predict future preferences with high accuracy. However, this process creates a "privacy paradox": the system requires granular personal data to function, yet centralizing this data exposes users to significant risks. If a recommendation server is compromised or the provider is untrustworthy, a user's entire behavioral profile becomes vulnerable.

Vulnerability Taxonomy

The literature identifies three primary threats to user privacy in CF systems:

  1. User Profile Exposure: The risk of raw user data being leaked or sold to third parties.
  2. Inference Attacks: Malicious actors using system outputs to deduce private preferences.
  3. Shilling Attacks: The injection of fake profiles to manipulate recommendations (e.g., boosting a product's rank).

Research in the field is often categorized by which of these specific threats it aims to mitigate. As shown in the classification table below, User Profile Exposure has historically received the most academic attention, though Inference Attacks represent a growing and more complex area of study.

Classification of PPCF literature by type of vulnerability addressed.

Technical Approaches to Privacy

Research in Privacy-Preserving Collaborative Filtering (PPCF) generally falls into four technical pillars:

  • Cryptography-based: Utilizing Homomorphic Encryption to perform mathematical operations on encrypted data, or Secure Multiparty Computation (SMC) for distributed data scenarios.
  • Obfuscation-based: Implementing Randomized Perturbation (adding noise) or k-Anonymity to mask individual identities while preserving aggregate statistics.
  • Clustering-based: Grouping users into communities and sharing only aggregate data, which balances identity protection with recommendation utility.
  • Hybrid & Domain-Specific: Leveraging item-based CF or expert opinions to reduce reliance on sensitive user profiles.

Key Insights and Trade-offs

A central theme in PPCF is the "Privacy-Accuracy-Performance" triangle. Cryptographic methods offer robust privacy but suffer from high computational overhead, making them difficult to scale for millions of users. Conversely, obfuscation techniques are efficient but introduce noise that degrades recommendation accuracy.

One notable observation is that item-based CF is inherently more privacy-friendly than user-based approaches. By focusing on item relationships rather than user similarities, systems can often operate with significantly less exposure of individual profiles. Furthermore, as systems move to the cloud, new challenges regarding "trusted third parties" emerge.

Conclusion

While User Profile Exposure is the most extensively researched area, Inference Attacks represent a stealthier threat requiring more sophisticated solutions like k-anonymity. Currently, clustering remains a practical "middle ground" for production systems, offering a balance of anonymity and computational efficiency. Future work must address the challenges of "malicious" actor models—moving beyond the standard "semi-honest" assumption—and the need for real-time updates in high-velocity data environments without compromising the delicate balance of privacy and performance.

Summary updated

Abstract

Collaborative filtering recommendation systems provide recommendations to users based on their own past preferences, as well as those of other users who share similar interests. The use of recommendation systems has grown widely in recent years, helping people choose which movies to watch, books to read, and items to buy. However, users are often concerned about their privacy when using such systems, and many users are reluctant to provide accurate information to most online services. Privacy-preserving collaborative filtering recommendation systems aim to provide users with accurate recommendations while maintaining certain guarantees about the privacy of their data. This survey examines the recent literature in privacy-preserving collaborative filtering, providing a broad perspective of the field and classifying the key contributions in the literature using two different criteria: the type of vulnerability they address and the type of approach they use to solve it.
BibTeX
@misc{elnabarawy2020survey,
  author = {Islam Elnabarawy and Wei Jiang and Donald C. Wunsch},
  title = {{Survey of Privacy-Preserving Collaborative Filtering}},
  year = {2020},
  eprint = {2003.08343},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2003.08343},
}
Figure