Multi-Party Data Release with Gaussian Noise for Privacy-Utility Balance
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing differentially private (DP) data release techniques are inadequate for multi-party settings where multiple parties each own disjoint datasets of the same data subjects, leading to reduced utility in machine learning models and privacy concerns.
Innovation Solution
Two improved DP algorithms are introduced: De-biased Gaussian Mechanism for Ordinary Least Squares (DGM-OLS) and Random Mixing Gaussian Mechanism for Ordinary Least Squares (RMGM-OLS), which utilize Gaussian noise and a shared Bernoulli projection matrix to ensure differential privacy while maintaining data utility.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If existing differentially private data release techniques are used in multi-party settings, then individual privacy is protected, but data utility for machine learning models is reduced
Solution Approach 1:
The patent segments the data release process into multiple independent steps: (1) each party independently adds Gaussian noise to their local data, (2) noisy data is aggregated by a central server, (3) a global model is trained on the aggregated noisy data. This segmentation allows each party to maintain control over their data while contributing to a collective model, preserving privacy without sacrificing overall data utility.
Solution Approach 2:
The patent combines noisy data from multiple independent parties into a single aggregated dataset that is then used for global model training. By merging the noisy contributions from all parties, the system achieves sufficient statistical power for accurate machine learning while each individual's contribution remains protected by the noise added at the source.
2Loss of information
If multiple parties share their datasets publicly, then data utility for machine learning is improved, but individual privacy is compromised
Solution Approach 1:
The patent applies preliminary action by having each party add Gaussian noise to their data BEFORE sharing it with others. This pre-processing step ensures that privacy protection is built into the data itself before aggregation, rather than being applied as a post-processing constraint that would reduce utility.
Solution Approach 2:
The patent introduces a central aggregation server as an intermediary that collects noisy data from all parties and distributes the global model updates. This intermediary enables coordinated multi-party learning without requiring any party to directly access or share their raw data with others, thus maintaining privacy while achieving data utility.
3Reliability
If noisy data from multiple parties is aggregated, then differential privacy is maintained, but communication overhead increases
Solution Approach 1:
The patent extracts the privacy-protection mechanism (Gaussian noise addition) from the central aggregation process and places it at each individual party's local system. This extraction eliminates the need for complex centralized differential privacy management and reduces communication overhead, as each party independently handles their own privacy protection before sharing.
Data Source
AI summary
The present disclosure describes techniques of releasing data while protecting individual privacy. A dataset may be compressed by applying a first random matrix. The dataset may be owned by a party among a plurality of parties and there may be a plurality of datasets owned by the plurality of parties. A noise may be added by applying a random Gaussian matrix to the compressed dataset to obtain a processed dataset. The processed dataset ensures data privacy protection. The processed dataset may be released to other parties.