[Paper Review] Outis: Crypto-Assisted Differential Privacy on Untrusted Servers

Amrita Roy Chowdhury, Chenghong Wang+3 more|arXiv (Cornell University)|Feb 20, 2019

Privacy-Preserving Technologies in Data22 citations

TL;DR

Cryptε proposes a crypto-assisted differential privacy system that achieves central-model accuracy and expressivity without requiring a trusted data collector by using two non-colluding untrusted servers to compute on encrypted data. It enables secure, efficient DP analysis through automated compilation of logical DP programs into privacy-preserving secure protocols with strong security guarantees.

ABSTRACT

Differential privacy (DP) has steadily become the de-facto standard for achieving privacy in data analysis, which is typically implemented either in the or model. The local model has been more popular for commercial deployments as it does not require a trusted data collector. This increased privacy, however, comes at a cost of utility and algorithmic expressibility as compared to the central model. In this work, we propose, Crypt$\epsilon$, a system and programming framework that (1) achieves the accuracy guarantees and algorithmic expressibility of the central model (2) without any trusted data collector like in the local model. Crypt$\epsilon$ achieves the best of both worlds by employing two non-colluding untrusted servers that run DP programs on encrypted data from the data owners. Although straightforward implementations of DP programs using secure computation tools can achieve the above goal theoretically, in practice they are beset with many challenges such as poor performance and tricky security proofs. To this end, Crypt$\epsilon$ allows data analysts to author logical DP programs that are automatically translated to secure protocols that work on encrypted data. These protocols ensure that the untrusted servers learn nothing more than the noisy outputs, thereby guaranteeing DP (for computationally bounded adversaries) for all Crypt$\epsilon$ programs. Crypt$\epsilon$ supports a rich class of DP programs that can be expressed via a small set of transformation and measurement operators followed by arbitrary post-processing. Further, we propose performance optimizations leveraging the fact that the output is noisy. We demonstrate Crypt$\epsilon$'s feasibility for practical DP analysis with extensive empirical evaluations on real datasets.

Motivation & Objective

• To bridge the gap between the high utility of the central model and the strong trust guarantees of the local model in differential privacy.
• To enable practical, efficient, and secure execution of DP programs on untrusted servers without requiring a trusted data collector.
• To support a rich class of DP workloads through a high-level programming framework that compiles to secure computation protocols.
• To address performance and security proof challenges in deploying DP via secure computation on encrypted data.
• To demonstrate feasibility and efficiency of the approach on real-world datasets through extensive empirical evaluation.

Proposed method

• Cryptε uses two non-colluding untrusted servers to jointly compute on encrypted data from data owners, ensuring no server learns more than the final noisy output.
• It compiles high-level logical DP programs—expressed via transformation and measurement operators—into secure computation protocols that preserve differential privacy.
• The system leverages homomorphic encryption and secure multi-party computation techniques to enable computation on encrypted data while maintaining privacy for computationally bounded adversaries.
• It introduces performance optimizations that exploit the fact that the output of DP programs is inherently noisy, reducing redundant computation and communication.
• The framework ensures end-to-end differential privacy by guaranteeing that untrusted servers learn only the noisy output, not intermediate or raw data.
• It supports arbitrary post-processing on the final output, preserving the privacy guarantees of the original computation.

Experimental results

Research questions

• RQ1 Can we achieve the accuracy and expressivity of the central model in differential privacy without requiring a trusted data collector?
• RQ2 Can secure computation on encrypted data be made efficient and practical for real-world DP workloads?
• RQ3 How can we ensure strong privacy guarantees (for computationally bounded adversaries) in a system with untrusted servers?
• RQ4 What optimizations are possible when the output of a computation is inherently noisy, as in DP?
• RQ5 Can a high-level programming framework for DP be automatically compiled into secure, privacy-preserving protocols?

Key findings

• Cryptε successfully achieves the accuracy and algorithmic expressivity of the central model while maintaining the trustworthiness of the local model by eliminating the need for a trusted data collector.
• The system demonstrates practical performance on real datasets, showing that secure computation for DP can be efficient enough for real-world deployment.
• The framework supports a broad class of DP programs through a small set of transformation and measurement operators, enabling expressive yet secure analysis.
• Performance optimizations based on the noise inherent in DP outputs significantly reduce communication and computation overhead.
• The security model ensures that untrusted servers learn nothing beyond the final noisy output, providing formal differential privacy guarantees for computationally bounded adversaries.
• Empirical evaluations confirm that Cryptε maintains strong privacy while delivering competitive performance compared to traditional DP implementations.