Cross-Company Data Aggregation With Relevance Selection and Privacy Budget
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing data utilization among multiple companies lacks examination on which data items should be linked to improve utility of statistical information, and the number of combinations is vast, making it impractical to try aggregation for all combinations due to privacy budget limitations.
Innovation Solution
An information processing device that calculates relevance between an index related to data utilization and data items using supervised machine learning, selects data items for aggregation patterns, and adds noise based on differential privacy criteria within a privacy budget.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of information
If aggregation is executed for all combinations of data items among multiple companies, then the utility of statistical information is improved, but the privacy budget is exhausted and computational complexity becomes unmanageable
Solution Approach 1:
The patent applies preliminary action by calculating relevance between data items and the analysis target before executing aggregation. The relevance calculation unit computes relevance scores for all candidate data items based on their relationship to the analysis target, and the selection unit pre-selects high-relevance data items before aggregation is performed. This preliminary selection process reduces the number of aggregation operations needed while ensuring that the most useful data items are included, thereby improving statistical information utility without exhausting computational resources or privacy budget.
Solution Approach 2:
The patent applies the extraction principle by isolating and selecting only the essential data items that have high relevance to the analysis target. Rather than processing all possible data item combinations, the selection unit extracts the subset of data items that contribute most significantly to the statistical analysis. This extraction approach maintains high utility of statistical information while dramatically reducing computational complexity and privacy budget consumption by focusing only on the most relevant data.
2Reliability
If noise is added to aggregation results to protect privacy, then privacy protection is improved, but the accuracy of statistical information deteriorates
Solution Approach 1:
The patent applies partial action by adding noise only to the aggregation results of selected high-relevance data items rather than to all possible data item combinations. The selection unit identifies a limited subset of data items with relevance scores above a threshold, and noise is added only to these selected items. This partial approach maintains adequate privacy protection for the most critical data while preserving higher accuracy in the statistical information compared to adding noise to all combinations.
3Loss of energy
If data items are selected based on relevance calculation, then the privacy budget is preserved, but the utility of statistical information may be reduced if important data items are excluded
Solution Approach 1:
The patent applies feedback by using the relevance calculation results to guide the selection of data items for aggregation. The relevance calculation unit computes relevance scores based on the relationship between each data item and the analysis target, providing feedback that informs the selection unit's decisions. This feedback mechanism ensures that data items with higher relevance to the analysis target are prioritized for inclusion in the aggregation, thereby preserving privacy budget by limiting the number of aggregations while maintaining the utility of statistical information through selective inclusion of important data items.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
An information processing device (100) that, when executing aggregation targeting an aggregation pattern including multiple combinations of data items contained in attribute information between the information processing device (100) itself and a counterpart device, both holding user data including a user ID and the attribute information related to a user, adds noise based on differential privacy criteria to aggregation results within a privacy budget, and executes the aggregation targeting the aggregation pattern, including: a relevance calculation unit (101) that calculates relevance between an index related to a purpose of data utilization and the data item contained in the attribute information; and a selection unit (102) that selects data items to be included in the aggregation pattern based on the calculated relevance.