DPText Privacy Preserving Text Representation Learning Framework
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing privacy-preserving techniques for user-generated textual data are inefficient due to its unstructured and noisy nature, often resulting in significant utility loss while failing to effectively anonymize and protect semantic meaning and private attributes.
Innovation Solution
A double privacy preserving text representation learning framework (DPText) that utilizes an auto-encoder, differential-privacy-based noise adder, semantic meaning discriminator, and private attribute discriminator to learn a modified latent representation that is differentially private, obscures private attributes, and retains semantic meaning by adding optimal noise to the text representation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional privacy preserving techniques (k-anonymity, differential privacy) are applied to user-generated textual data, then privacy protection is improved, but utility is significantly lost due to the unstructured and noisy nature of the data
Solution Approach 1:
The patent transforms the unstructured textual data into a structured latent representation space where privacy can be effectively protected. By changing the parameter space from raw text to learned embeddings, the system achieves both privacy protection and utility preservation simultaneously, overcoming the limitation of traditional techniques that work poorly on unstructured data
Solution Approach 2:
The patent introduces an autoencoder as an intermediary component that learns to compress and reconstruct textual data. This intermediary transformation enables privacy-preserving processing by operating on the compressed latent representation rather than the original unstructured text, thereby maintaining both privacy and utility
2Reliability
If traditional anonymization techniques are used on textual data, then privacy is protected, but semantic meaning is lost due to the unstructured and noisy characteristics of the data
Solution Approach 1:
The patent changes the representation parameters of textual data from raw unstructured text to structured latent embeddings through an autoencoder. This parameter transformation preserves semantic meaning in the compressed space while enabling effective privacy protection, addressing the semantic loss problem of traditional anonymization
Solution Approach 2:
The patent replaces traditional mechanical anonymization approaches (such as token suppression or generalization) with a learned neural network-based transformation. This substitution allows the system to preserve semantic meaning by learning meaningful representations that maintain task utility while protecting privacy
3Reliability
If noise is added to text representation for privacy protection, then privacy is improved, but semantic meaning may be degraded
Solution Approach 1:
The patent applies differential privacy noise selectively and controlledly to the latent representation rather than the original text. By applying noise only to specific dimensions or with controlled magnitude in the latent space, the system achieves privacy protection while minimizing semantic degradation through the robustness of the learned representation
Data Source
AI summary
Various embodiments of a computer-implemented system which learns textual representations while filtering out potentially personally identifying data and retaining semantic meaning within the textual representations are disclosed herein.