Dimension-Scaled Noise Addition for N-Dimensional Text Embeddings
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing noise addition techniques for differential privacy reduce the utility of data for downstream use and can be reversed, compromising the privacy of sensitive information in data loss prevention (DLP) systems.
Innovation Solution
A noise addition service that scales noise per dimension based on localized minimum and maximum values for N-dimensional text embeddings, using techniques like the Laplace mechanism, to preserve data utility and obscure sensitive information without compromising classification accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If conventional global scaling noise addition is applied to N-dimensional embeddings, then privacy protection is improved, but data utility and classification accuracy deteriorate
Solution Approach 1:
The patent applies local quality by transitioning from global scaling to dimension-specific scaling. Each dimension of the N-dimensional embedding is scaled independently using its own minimum and maximum values, allowing the noise addition to preserve local data structures and relationships. This localized approach maintains classification accuracy while still providing privacy protection through differential privacy.
Solution Approach 2:
The patent changes the scaling parameters from a single global scale factor to multiple dimension-specific scale factors. By computing minimum and maximum values for each dimension separately and using these as scaling parameters, the system optimizes the noise addition process to preserve more information while maintaining privacy guarantees.
2Object-affected harmful factors
If noise is added to data for differential privacy, then the ability to infer sensitive information is reduced, but the utility of data for downstream analysis is reduced
Solution Approach 1:
By applying dimension-specific scaling before noise addition, the patent preserves local data characteristics and relationships that are critical for downstream analysis. This localized approach ensures that the most important features are less degraded by noise, maintaining data utility while still providing privacy protection.
Solution Approach 2:
The patent segments the N-dimensional embedding into individual dimensions and processes each dimension separately with its own scaling parameters. This segmentation allows the noise addition to be optimized for each dimension's characteristics, reducing overall information loss while maintaining privacy.
Data Source
AI summary
Noise is added to data obtained from customers for differential privacy without reducing utility of the data for downstream use and/or analysis, such as data obtained from data loss prevention (DLP) services that are used for ongoing learning of DLP models. Noise is added to an N-dimensional text embedding(s) based on scaling values contained in the text embeddings on a per-dimension basis. For each dimension of the embedding(s), the corresponding value at that dimension is scaled based on minimum and maximum values that are localized to that dimension and were previously selected based on experimental data for which embeddings were generated. Noise is added to the resulting embeddings that have been scaled per dimension, such as with the Laplace mechanism.