Differential Privacy via Transform Decorrelation for User Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current differential privacy mechanisms are vulnerable to inference attacks due to the assumption of independent data points, which is not reflective of real-world inter-user correlations, allowing attackers to infer individual identities more easily.
Innovation Solution
Applying a transform to decorrelate user data points and then using an (ε, δ)-differential privacy mechanism on the transform coefficients to create a private dataset, which is (ε, δ)-differentially private, thereby protecting user data from inference attacks while preserving fundamental information.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If differential privacy mechanisms are applied to data points assuming mutual independence, then calculations are simplified and privacy protection is provided, but attackers can exploit inter-user correlations to infer individual identities more effectively
Solution Approach 1:
The patent segments the data processing into two distinct stages: first applying a transform to decorrelate data points, then applying differential privacy mechanisms to the transformed coefficients. This segmentation allows the system to handle inter-user correlations explicitly while maintaining the simplicity of DP calculations on the transformed domain.
Solution Approach 2:
The patent applies a transform to the data points before applying the differential privacy mechanism. This preliminary action decorrelates the data points, removing inter-user correlations that would otherwise compromise privacy protection. The transformed coefficients are then processed by the DP mechanism, ensuring privacy guarantees hold even when attackers know correlation structures.
2Loss of information
If differential privacy mechanisms add independent noise to data outcomes, then fundamental information is retained while individual identities are hidden, but the mechanism fails to account for known inter-user correlations allowing enhanced inference attacks
Solution Approach 1:
The patent applies a transform to the data points before adding noise through the differential privacy mechanism. This preliminary transformation decorrelates the data, ensuring that the subsequent noise addition cannot be exploited by attackers who know inter-user correlations. The transformed coefficients maintain fundamental information while the noise ensures privacy.
Solution Approach 2:
The transform acts as an intermediary between the original correlated data and the differential privacy mechanism. It converts correlated data points into uncorrelated transform coefficients, which then serve as the input for the DP mechanism. This intermediary step isolates the DP mechanism from the correlation structure, preventing attackers from exploiting correlations to reduce the effectiveness of the noise.
3Reliability
If data points are transformed to decorrelate them before applying differential privacy, then privacy protection against inference attacks is enhanced, but the complexity of the processing system increases
Solution Approach 1:
The patent changes the parameter space by applying a transform to convert correlated data points into uncorrelated transform coefficients. This parameter transformation simplifies the subsequent differential privacy processing by eliminating inter-user correlations, making the system more efficient rather than less complex.
Solution Approach 2:
The patent replaces the complex task of handling correlated data directly with a simpler process: transform the data to remove correlations, then apply standard DP mechanisms. This substitution of the mechanical approach (handling correlations directly) with a mathematical transformation approach simplifies the overall processing while enhancing privacy protection.
Data Source
AI summary
A method for increasing privacy of user data of a plurality of users within a dataset is disclosed. The method comprises, in one or more data processing devices, providing (10) a dataset comprising a plurality of data points of a plurality of users and comprising inter-user correlations within the plurality of data points; determining (12) a plurality of transform coefficients by applying a transform on the plurality of data points; determining (14) a plurality of private transform coefficients from the plurality of transform coefficients by applying an (ε, δ)- differential privacy mechanism to each non-zero transform coefficient of the plurality of transform coefficients; and determining (15) a private dataset comprising a plurality of private data points from the plurality of private transform coefficients by applying, on the plurality of private transform coefficients, an inverse transform of the transform; wherein the (ε, δ)- differential privacy mechanism is adapted such that the plurality of private data points is (ε, δ)-differential private. Further, a system for increasing privacy of user data of a plurality of users within a dataset and a computer program product are provided.