RLTA Text Anonymization via Embedding Perturbation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for anonymizing textual user information fail to balance privacy and utility, often leading to privacy breaches and loss of data quality, as they do not effectively obscure private attributes like age, gender, and location while preserving the semantic meaning for machine learning tasks.
Innovation Solution
A deep reinforcement learning-based system, Reinforcement Learning Text Anonymizer (RLTA), uses a Markov Decision Process to learn an optimal strategy for manipulating text embeddings, incorporating an attention-based task-aware text representation learner and a deep Q-learning privacy and utility preserver to obscure private attributes while maintaining data utility through a reward function that balances privacy and utility.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional anonymization methods are applied to textual data, then privacy protection is improved, but data utility and quality are degraded
Solution Approach 1:
The patent applies parameter changes by modifying text embeddings through controlled perturbations in the vector space. The reinforcement learning agent learns to adjust embedding parameters (dimensions and values) to achieve privacy protection while preserving data utility. This involves changing the numerical parameters of text representations rather than the original text, allowing fine-grained control over the privacy-utility tradeoff.
Solution Approach 2:
The patent introduces text embeddings as an intermediary representation between the original text and the anonymization process. Instead of directly anonymizing the original text, the system transforms text into embedding vectors, which then serve as the target for privacy-preserving manipulations. This intermediary layer allows privacy protection to be applied in a continuous vector space while maintaining the semantic relationships needed for data utility.
2Adaptability or versatility
If text embeddings are shared for different machine learning tasks, then adaptability is improved, but privacy leakage risk increases
Solution Approach 1:
The patent applies preliminary action by performing privacy-preserving transformations on text embeddings before they are shared or used for multiple machine learning tasks. The reinforcement learning agent learns and applies anonymization strategies in advance, ensuring that privacy protection is built into the embeddings before they are exposed to potential inference attacks. This proactive approach prevents privacy leakage while maintaining the embeddings' usefulness for various tasks.
Data Source
AI summary
Various embodiments of a system and associated method for anonymization of text without losing semantic utility of text by extracting a latent embedding representation of content with respect to a given task and by learning an optimal strategy for text embedding manipulation to satisfy both privacy and utility requirements are disclosed herein. In particular, the system balances private attribute obfuscation with retained semantic utility.