Towards Statistical Queries over Distributed Private User Data (NSDI '12)

2012-11-20 16:13

From MPI-SWS, Germany, paper and slides

  1. Problem

To protect user privacy in distributed systems from leaking by statistical queries.

  1. Challenges

The most direct solutions are

  1. to anonymize + add noise to user data.

    [-] utility, de-anonymize

  2. differential privacy. add noise to answer of queries.

    [-] scale, churn tolerance, malicious client

  3. Solution

PDDP: Practical Distributed Differential Privacy.

  • Binary answer in bucket. The query result should not be distorted by the client arbitrarily.
  • Blind noise addition. The Malicious should not be trusted. Private data are controlled by its user only.

3.1 Assumption

Clients and analysts are potentially malicious. Proxy is HbC (honest but curious) and should not have access to noise-free result.

3.2 Work Flow

  1. Query Initialization(Analyst -> Proxy)

  2. Query Forwarding (Proxy -> Client)

  3. Client Respond (Client -> Proxy)

    answers are encrypted with the analyst's public key.

  4. Differential Private Noise Addition.

    collaborative coin generation with a GM cryptosystem. Unbiased proxy flip encrypted coins from clients randomly and thus transform them into unbiased ones. Coins serve as DP noises.

  5. Noisy Answers to Analyst ( Proxy -> Analyst)

3.3 Implementation and Deployment

600+ Client = Firefox add-on + SQLite

Proxy = Tomcat web service + MySQL

Analyst = Java program

  1. Conclusion

The authors achieve scalable, churn-tolerant user privacy against malicious analyst and clients by

  1. making a trade off between utility and privacy. (differential privacy)
  2. introduce distributed system to traditionally centralized differential privacy environment. (distributed)
© 2010-2018 Tian
Built with ❤️ in San Francisco