[Paper Review] I Am Not What I Write: Privacy Preserving Text Representation Learning

Ghazaleh Beigi, Kai Shu+3 more|arXiv (Cornell University)|Jul 6, 2019

Privacy-Preserving Technologies in Data44 references18 citations

TL;DR

This paper proposes DPText, a novel framework for learning privacy-preserving text representations that simultaneously ensure differential privacy, obscure private attributes, and retain high utility for downstream NLP tasks. By combining an autoencoder with differentially private noise, a semantic discriminator, and a private attribute discriminator, DPText effectively reduces re-identification risk and private information leakage while preserving semantic meaning, as validated on sentiment analysis and part-of-speech tagging tasks.

ABSTRACT

Online users generate tremendous amounts of textual information by participating in different activities, such as writing reviews and sharing tweets. This textual data provides opportunities for researchers and business partners to study and understand individuals. However, this user-generated textual data not only can reveal the identity of the user but also may contain individual's private information (e.g., age, location, gender). Hence, "you are what you write" as the saying goes. Publishing the textual data thus compromises the privacy of individuals who provided it. The need arises for data publishers to protect people's privacy by anonymizing the data before publishing it. It is challenging to design effective anonymization techniques for textual information which minimizes the chances of re-identification and does not contain users' sensitive information (high privacy) while retaining the semantic meaning of the data for given tasks (high utility). In this paper, we study this problem and propose a novel double privacy preserving text representation learning framework, DPText, which learns a textual representation that (1) is differentially private, (2) does not contain private information and (3) retains high utility for the given task. Evaluating on two natural language processing tasks, i.e., sentiment analysis and part of speech tagging, we show the effectiveness of this approach in terms of preserving both privacy and utility.

Motivation & Objective

• To address the challenge of protecting user privacy in user-generated textual data, which often reveals identities and sensitive attributes.
• To design a method that preserves the semantic utility of text for downstream NLP tasks while minimizing re-identification and attribute inference risks.
• To overcome limitations of existing techniques such as PII removal, k-anonymity, and standard differential privacy, which fail to protect against inference attacks on latent representations.
• To develop a unified framework that integrates differential privacy with representation learning to simultaneously protect privacy and maintain utility.
• To ensure that learned text representations are both semantically meaningful and resistant to private attribute inference and re-identification.

Proposed method

• The framework employs an autoencoder to learn compressed, low-dimensional representations from raw text input.
• Differential privacy is enforced by adding Laplacian noise to the latent representation, ensuring that individual data points cannot be re-identified.
• A semantic meaning discriminator is trained to preserve the original text's semantic content, ensuring utility for downstream tasks.
• A private attribute discriminator is used to minimize the model's ability to infer sensitive attributes (e.g., age, gender, health conditions) from the representation.
• The model is trained end-to-end using adversarial training: the autoencoder learns to generate representations that fool the private attribute discriminator while satisfying the semantic discriminator and differential privacy constraints.
• The framework is agnostic to the initial representation model, allowing integration with existing methods such as doc2vec.

Experimental results

Research questions

• RQ1 Can a text representation learning framework simultaneously ensure differential privacy, obscure private attributes, and preserve semantic utility?
• RQ2 How effective is the proposed framework in reducing the risk of re-identification of individual text representations?
• RQ3 To what extent does the framework prevent inference of private attributes such as gender, age, or medical conditions from learned representations?
• RQ4 How does the framework compare to existing methods in terms of utility and privacy protection on real-world NLP tasks?
• RQ5 Does the integration of semantic and private attribute discriminators improve privacy-utility trade-offs compared to standard differential privacy or adversarial training alone?

Key findings

• DPText significantly reduces the risk of re-identification of individual text representations by enforcing differential privacy through Laplacian noise injection in the latent space.
• The framework successfully obscures private attribute information, as demonstrated by the private attribute discriminator's inability to accurately predict sensitive attributes from the learned representations.
• The semantic meaning of the original text is well-preserved, as evidenced by strong performance on downstream tasks such as sentiment analysis and part-of-speech tagging.
• Compared to prior work, DPText outperforms methods that rely solely on PII removal or standard differential privacy, which fail to protect against inference from latent representations.
• The adversarial training setup with dual discriminators ensures a better privacy-utility trade-off than existing approaches that do not explicitly model both semantic fidelity and private attribute leakage.
• The framework is robust and generalizable, as it does not depend on a specific pre-trained representation model and can be integrated with existing text embedding techniques.