Differentially Private Oblivious RAM Sameer Wagh , Paul Cuff , - PowerPoint PPT Presentation

Differentially Private Oblivious RAM Sameer Wagh ∗ , Paul Cuff † , Prateek Mittal ∗ July 24, 2019 ∗ Princeton University, † Renaissance Technologies

Introduction: Oblivious RAM Access data privately from private database. 1

Introduction: Oblivious RAM User receives record R 2

Introduction: Oblivious RAM Obliviousness: Adversary should not know R 3

ORAM Application I Client-server environments 4

ORAM Application II Trusted Execution Environments such as SGX-based enclaves 5

The Problem? 6

The Problem? 7

The Problem: Overhead Overheads • Logarithmic bandwidth overhead ( ≥ 100 × ) • Logarithmic storage overhead 8

Key Insight Overheads • Logarithmic bandwidth overhead ( ≥ 100 × ) • Logarithmic storage overhead Key Insight Can we improve performance by relaxing privacy? 9

Key Insight: Improve Performance by Relaxing Privacy • Statistically private ORAM ◮ Better performance at the cost of privacy loss ◮ Challenge: Can we provide rigorous guarantees? 10

Key Insight: Improve Performance by Relaxing Privacy • Statistically private ORAM ◮ Better performance at the cost of privacy loss ◮ Challenge: Can we provide rigorous guarantees? • Efficiency ◮ Reduce performance overheads – bandwidth, local storage ◮ Achieve privacy proportional to application resources 10

Key Insight: Improve Performance by Relaxing Privacy Differential Privacy • Statistically private ORAM ◮ Better performance at the cost of privacy loss • Formalize Differentially Private ORAM ◮ Challenge: Can we provide rigorous guarantees? • Introduce Root ORAM • Efficiency ◮ Reduce performance overheads – bandwidth, local storage ◮ Achieve privacy proportional to application resources 11

Key Insight: Improve Performance by Relaxing Privacy Differential Privacy • Statistically private ORAM ◮ Better performance at the cost of privacy loss • Formalize Differentially Private ORAM ◮ Challenge: Can we provide rigorous guarantees? • Introduce Root ORAM Root ORAM • Efficiency ◮ Reduce performance overheads – bandwidth, local storage • Theoretical Results ◮ Achieve privacy proportional to application resources • Empirical Results • Private Information Retrieval 12

Differentially Private Oblivious RAM

DP-ORAM Intuition 13

Statistical closeness - Differential Privacy Pr [ ORAM ( a 1 ) ∈ S ] ≤ e ǫ Pr [ ORAM ( a 2 ) ∈ S ] + δ 16

Protocol Construction

Root ORAM: Storage 17

Root ORAM: Invariant 18

Root ORAM: Updated mapping 19

Root ORAM: Updated mapping Key Insight • Uniform mapping ⇒ Conventional Security • Non-uniform mapping ⇒ DP-ORAM Security 21

Root ORAM: Updated mapping Key Insight • Uniform mapping ⇒ Conventional Security • Non-uniform mapping ⇒ DP-ORAM Security 22

Root ORAM: Non-Uniform mapping Impact • Lower average placement ⇒ Improved performance • Privacy loss 24

Root ORAM: Write back 25

Root ORAM: Lowest Common Intersection 26

Database view before access 29

Database view after access 30

Results

Security Result: Root ORAM is DP-ORAM Differentially Private ORAM Protocol The Root ORAM protocol with parameters k , p is ( ǫ, δ ) -differentially private for the following choice of ǫ and δ � 1 + (2 k − 1) · p � ǫ = 2 log 1 − (1 − δ k 0 ) p (1) � M � 1 + (2 k − 1) · p δ = M · N where δ k 0 is the Kronecker delta, M is the size of the access sequence and M > total stash size. 31

Performance Improvements Improvement in stash usage for ( L , k , Z ) = (15 , 1 , 4) 32

Performance Improvements Key takeaway DP-ORAM can enhance performance at the cost of privacy Improvement in stash usage for ( L , k , Z ) = (15 , 1 , 4) 33

Application: Private Information Retrieval

Private Information Retrieval (PIR) Access data privately from public database. [46] Mittal, Prateek, Femi G. Olumofin, Carmela Troncoso, Nikita Borisov, and Ian Goldberg. ”PIR-Tor: Scalable Anonymous Communication Using Private Information Retrieval.” In USENIX Security Symposium , p. 31. 2011. 34

ORAM based PIR • ORAM has been used previously for PIR [7, 59] [7] Michael Backes, Aniket Kate, Matteo Maffei, and Kim Pecina. ObliviAd: Provably secure and practical online behavioral advertising. In IEEE Symposium on Security and Privacy (S&P) , 2012. [59] Peter Williams and Radu Sion. Usable PIR. In Symposium on Network and Distributed System Security (NDSS) , 2008. 35

DP-PIR Bandwidth Comparison Security-Bandwidth trade-offs for DP-PIR protocols (Toledo et.al. [54], Path-PIR [42], and Path ORAM [53]). 41

DP-PIR Bandwidth Comparison DP-ORAMs provide significant performance benefits for DP-PIR Security-Bandwidth trade-offs for DP-PIR protocols (Toledo et.al. [54], Path-PIR [42], and Path ORAM [53]). 42

Conclusion

Summary • Formalized Differentially Private ORAMs • Introduced a family of DP-ORAM protocols • Analyzed security, performance • Showcased utility for Private Information Retrieval 43

Summary • Formalized Differentially Private ORAMs • Introduced a family of DP-ORAM protocols • Analyzed security, performance • Showcased utility for Private Information Retrieval • Possible to enhance performance by relaxing privacy 43

Summary • Formalized Differentially Private ORAMs • Introduced a family of DP-ORAM protocols • Analyzed security, performance • Showcased utility for Private Information Retrieval • Possible to enhance performance by relaxing privacy Source code is available at https://github.com/inspire-group/Root-ORAM 43

Thank you! 43

Thank you! Questions? 43

Differentially Private Oblivious RAM Sameer Wagh , Paul Cuff , - PowerPoint PPT Presentation

Differentially Private Oblivious RAM Sameer Wagh , Paul Cuff , Prateek Mittal July 24, 2019 Princeton University, Renaissance Technologies Introduction: Oblivious RAM Access data privately from private database. 1

Estimating the Variance of Complex Differentially Private Algorithms Robert Ashmead JSM 2019,

Verifying Differentially Private Bayesian Inference Marco Gaboardi University of Dundee Joint

Development, Evaluation, and Management of Differentially Private ML Pipelines on

Differentially-Private Federated Linear Bandits Introduction Federated Learning Contextual

Differentially-Private Deep Learning from an Optimization Perspective Presenter: Liyao Xiang

Differentially Private Model Publishing for Deep Learning Lei Yu, Ling Liu, Calton Pu , Mehmet

Differentially Private Algorithm and Auction Configuration Ellen Vitercik CMU, Theory Lunch

Differentially Private Distributed Convex Optimization via Functional Perturbation Erfan Nozari

Lo Locally Differentially Private Frequency Es Esti timati tion on Ex Exploi oiti ting Con

Releasing a Differentially Private Password Frequency Corpus from 70 Million Yahoo! Passwords

Differentially Private Recommender Systems David Madras University of Toronto April 4, 2017

Differentially Private Markov Chain Monte Carlo o 2 , Onur Dikmen 3 and Antti Honkela 1 a 1

Towards Practical Differentially Private Convex Optimization ROGER JOSEPH P. DAWN SONG IYENGAR

A Predictive Differentially-Private Mechanism for Mobility Traces Marco Stronati

Formal Verification of Differentially Private Mechanisms Marco Gaboardi University at Buffalo,

Differentially-Private Network Trace Analysis Frank McSherry and Ratul Mahajan Microsoft

Differentially Private Testing of Identity and Closeness of Discrete Distributions NeurIPS 2018,

Building Blocks of Privacy: Differentially Private Mechanisms Graham Cormode graham@cormode.org

Efficient Private File Retrieval by Combining ORAM and PIR

FUNCTIONALLY OBLIVIOUS (AND SUCCINCT) Edward Kmett BUILDING BETTER TOOLS Cache-Oblivious

Verifiable Homomorphic Oblivious Transfer and Private Equality Test Helger Lipmaa Helsinki

Perfectly S Secure O Oblivious A s Algorithms s in t the M Multi-Server S Setting T-H.

Oblivious Routing on Geometric Networks Costas Busch, Malik Magdon-Ismail and Jing Xi {

Lower Bound! Kasper Green Larsen Jesper Buus Nielsen Oblivious RAM Introduced by Goldreich