Hyperbolic Embedding Redaction for Privacy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing data privacy solutions, such as regex pattern matching and anonymization, are non-private as they fail to account for side knowledge that attackers can use to identify individuals, and subtler privacy breaches occur when innocuous data is combined with external information, making it difficult to protect privacy effectively.
Innovation Solution
The use of hierarchical word embeddings in a Hyperbolic representation to determine hypernyms for potentially sensitive words, introducing noise to preserve privacy while maintaining the intent and utility of the text, employing differential privacy mechanisms to scramble text in a way that protects privacy and intent.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If differential privacy mechanisms are applied to scramble text, then privacy preservation is improved, but text utility and meaning may deteriorate
Solution Approach 1:
The patent changes the parameter space from Euclidean to Hyperbolic geometry for word embeddings. This parameter change allows noise to be added in a way that preserves hierarchical relationships and semantic meaning while still providing differential privacy guarantees. The Hyperbolic space parameters enable better preservation of text utility compared to traditional Euclidean approaches.
Solution Approach 2:
The patent transitions from 2D Euclidean space to n-dimensional Hyperbolic space for representing word embeddings. This dimensional change provides additional degrees of freedom for adding noise while maintaining semantic relationships. The higher-dimensional Hyperbolic space allows for more effective privacy preservation without completely destroying text utility.
2Reliability
If hierarchical word embeddings in Hyperbolic representation are used to determine hypernymis, then privacy preservation is improved, but computational complexity increases
Solution Approach 1:
The patent pre-computes and stores hierarchical relationships (hypernymis) in the Hyperbolic space before the privacy-preserving processing occurs. By establishing these hierarchical structures in advance, the system avoids complex real-time computations during the actual text processing and noise addition phases, thereby reducing overall computational complexity while maintaining strong privacy guarantees.
3Reliability
If noise is introduced to preserve privacy, then privacy protection is improved, but measurement precision of text data deteriorates
Solution Approach 1:
The patent changes the geometric parameters of the embedding space from Euclidean to Hyperbolic, which fundamentally alters how noise affects the data. In Hyperbolic space, the curvature parameter allows noise to be distributed in a way that preserves hierarchical structures and semantic relationships better than Euclidean space. This parameter change enables privacy protection through noise addition while maintaining better measurement precision of the underlying text data.
Data Source
AI summary
A hypernym of a word in utterance data may be probabilistically determined. The utterance data may correspond to a spoken query or command. A redacted utterance may be derived by replacing the word with the hypernym. The hypernym may be determined by applying noise to a position in a hierarchical embedding that corresponds to the word. The word may be identified as being potentially sensitive. The hierarchical embedding may be a Hyperbolic embedding that may indicate hierarchical relationships between individual words of a corpus of words, such as “red” is a “color” or “Austin” is in “Texas.” Noise may be applied by obtaining a first value in Euclidean space based on a second value in Hyperbolic space, and obtaining a third value in Hyperbolic space based on the first value in Euclidean space. The second value in Hyperbolic space may correspond to the word.