Differentially Private PCA for Vertically Partitioned Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current approaches to differential privacy in federated learning primarily focus on horizontally partitioned data, and there is a lack of efficient and trusted methods for vertically partitioned data, where data privacy is not adequately preserved, especially in vertically federated learning scenarios.
Innovation Solution
The implementation of differentially private principal component analysis (PCA) for vertically partitioned data using Skellam noise and the Ben-Or, Goldwasser, and Widgerson (BGW) secret sharing protocol, where clients discretize their column vectors, introduce noise, and combine results securely to protect data privacy without revealing private inputs.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If MPC is applied to FL to prevent privacy leakage during computation, then computation security is improved, but the outcome still leaks information about private inputs
Solution Approach 1:
The patent introduces differential privacy as an intermediary mechanism between MPC computation and the final outcome. By adding calibrated noise to the computation results, the differential privacy mechanism acts as a mediator that preserves the usefulness of the outcome while preventing information leakage about private inputs. This resolves the contradiction by layering differential privacy on top of MPC to address the remaining information leakage issue.
2Reliability
If DP is applied to horizontally partitioned data in FL, then privacy protection is improved, but no efficient approach exists for vertically partitioned data
Solution Approach 1:
The patent segments the vertically partitioned data processing into distinct phases: local computation phase where each client computes on their own vertical partition, aggregation phase where results are combined, and differential privacy phase where noise is added. This segmentation allows differential privacy to be effectively applied to vertical FL by breaking down the complex vertical data structure into manageable components that can be processed with existing DP techniques.
3Reliability
If clients keep data localized in FL, then data privacy is preserved, but secure computation requires additional protocols like MPC
Solution Approach 1:
The patent enables clients to perform differential privacy computations on their own localized data without requiring complex multi-party computation protocols. Each client independently adds differential privacy noise to their local computations and shares only the perturbed results with the server. This self-service approach maintains data privacy through localization while eliminating the need for complex MPC protocols, as the differential privacy mechanism provides the necessary security guarantees.
Data Source
AI summary
Provided herein are systems, methods, and computer-readable media for implementing differentially private principal component analysis (PCA) for vertically partitioned data. An example system can include a first client possessing a first column vector, and a second client possessing a second column vector. The first client can be configured to discretize the first column vector to obtain a first discretized column vector and the second client can be configured to discretize the second column vector to obtain a second discretized column vector. The first client can be configured to introduce a first noise to the first discretized column vector and the second client can be configured to introduce a second noise to the second discretized column vector to obtain a PCA result.