Secure Multi-Party Computation Using Partial Noise Segmentation
Solution Overview
Problem
Existing secure multi-party computation methods for differential privacy are inefficient, leading to high communication and computation overhead, and require many evaluations of expensive exponential functions, making them unsuitable for secure data processing in multiparty environments.
Innovation Solution
The implementation of an exponential mechanism using noise from distributions like the Gumbel or exponential distribution, where each client provides partial noise terms that are securely combined to determine a noisy utility score for privacy-preserving statistics, allowing for efficient computation of differential privacy mechanisms without revealing individual data sets.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional secure multi-party computation methods are used for differential privacy, then privacy preservation is achieved, but computational efficiency deteriorates due to high communication and computation overhead
Solution Approach 1:
The patent segments the noise generation process into independent partial noise terms that can be computed separately by different parties or systems. Instead of computing a single large exponential function, the system divides the computation into smaller partial noise values that are subsequently summed to achieve the same differential privacy guarantee, thereby reducing computational overhead and improving efficiency.
2Reliability
If exponential functions are evaluated frequently for differential privacy, then privacy guarantees are maintained, but computation time increases significantly
Solution Approach 1:
The patent applies preliminary action by pre-computing or independently generating partial noise terms before the final aggregation step. Each party can compute their partial noise contribution in advance using simpler operations, and these pre-computed values are then combined to form the final noisy output. This eliminates the need for multiple expensive exponential function evaluations during the main computation process.
3Measurement precision
If plaintext data is shared for analytics, then computation accuracy is improved, but data security is compromised
Solution Approach 1:
The patent introduces noise as an intermediary element between the raw data and the analytics computation. Instead of sharing plaintext data directly, the system adds controlled random noise to the data or to intermediate computation results. This intermediary noise layer preserves the statistical properties needed for accurate analytics while preventing the exposure of sensitive individual data points, thus maintaining both accuracy and security.
Data Source
AI summary
Aspects of the current subject matter are directed to performing privacy-preserving analytics over sensitive data without sharing plaintext data. According to an aspect, a system includes at least one data processor and at least one memory storing instructions which, when executed by the at least one data processor, result in operations including: receiving, from each of a plurality of clients, a utility score and a partial noise value; performing, based on the received utility scores and the partial noise values, a secure multi-party computation of a privacy-preserving statistic, the performing of the secure multi-party computation of the privacy-preserving statistic further comprising determining a noisy utility score for each data value in a domain of output values and selecting a highest noise utility score from the determined noisy utilities scores; and providing, based on the selected highest utility score, an output value for the privacy-preserving statistic.