Differential Privacy Composition via Transformed Partial Convolution
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing differential privacy composition determination methods provide loose upper-bounds or introduce errors due to periodicity and truncation, leading to impractical implementation and inaccurate privacy loss quantification in multiple interactive queries.
Innovation Solution
A transformed partial convolution algorithm is used to determine privacy loss distribution, reducing computational complexity without introducing errors, and providing accurate privacy loss/leakage probability for differential privacy composition.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a full convolution operation is performed on the privacy loss distribution, then accurate privacy loss determination is achieved, but computational complexity increases significantly
Solution Approach 1:
The privacy loss distribution is segmented into multiple intervals based on privacy loss values. The convolution operation is then performed separately on each interval rather than on the entire distribution at once. This segmentation reduces the computational burden while maintaining accuracy within each segment, effectively resolving the contradiction between accurate privacy loss determination and computational complexity.
2Productivity
If Fast Fourier Transformation is used to reduce computation complexity, then computational speed improves, but computation errors are introduced
Solution Approach 1:
The patent extracts and removes the error-prone Fast Fourier Transformation step from the computation process. Instead, it directly performs interval-based convolution operations on the privacy loss distribution, taking out the transformation to frequency domain and multiplication steps that introduce errors. This extraction maintains computational reliability while still achieving reduced complexity through interval segmentation.
3Device complexity
If periodicity and truncation are applied to the privacy loss distribution, then computation complexity is reduced, but determination accuracy deteriorates due to introduced errors
Solution Approach 1:
The patent applies local quality by performing convolution operations with different levels of detail in different intervals of the privacy loss distribution. In intervals where high precision is critical, the computation maintains full accuracy, while in intervals where privacy loss is less significant, the computation can be simplified. This local differentiation reduces overall computational complexity without sacrificing accuracy where it matters most.
Data Source
AI summary
Differential privacy composition determination in a secure communication is provided. A method for determining differential privacy composition includes determining a differential privacy configuration including a first privacy parameter and a second privacy parameter, determining a privacy loss distribution, and performing a partial convolution operation of the privacy loss distribution by transforming the privacy loss distribution based on a predetermined parameter, determining an integral range based on the first privacy parameter and the privacy loss distribution, transforming the privacy loss distribution based on the integral range, and performing the partial convolution operation based on the integral range. The method also includes determining the second privacy parameter based on a result of the partial convolution operation, and controlling a dataset based on the differential privacy configuration to limit access to the dataset.