Privacy Protection in Conversations

Abstract

Leakage of personal information in online conversations raises serious privacy concerns. For example, malicious users might collect sensitive personal information from vulnerable users via deliberately designed conversations. This thesis tackles the problem of privacy leakage in textual conversations and proposes to mitigate the risks of privacy disclosure by detecting and rewriting the risky utterances. Previous research on privacy protection in text has a focus on manipulating the implicit semantic representations in a continuous high dimensional space, which are mostly used for eliminating trails of personal information to machine learning models. Our research has a focus on the explicit expressions of conversations, namely sequences of words or tokens, which are generally used between human interlocutors or human-computer interactions. The new setting for privacy protection in text could be applied to the conversations by individual human users, such as vulnerable people, and artificial conversational bots, such as digital personal assistants.

This thesis consists of two parts, answering two research questions: How to detect the utterances with the risk of privacy leakage? and How to modify or rewrite the utterances into the ones with less private information? In the first part of this thesis, we aim to detect the utterances with privacy leakage risk and report the sensitive utterances to authorized users for approval. One of the essential challenge of the detection task is that we cannot acquire a large-scale aligned corpus for supervised training of natural language inference for private information. A compact dataset is collect to merely validate the privacy leakage detection models. We investigate weakly supervised methods to learn utterance-level inference from coarse set-level alignment signals. Then, we propose novel alignment models for utterance inference. Our approaches manage to outperform competitive baseline alignment methods. Additionally, we develop a privacy-leakage detection system integrated in Facebook Messenger to demonstrate the utility of our proposed task in real-world usage scenarios. In the second part of this thesis, we investigate two pieces of work to rewrite the privacy-leakage sentences automatically into less sensitive ones. The first work discusses obscuring personal information in form of classifiable attributes. We propose to reduce the bias of sensitive attributes, such as gender, political slant and race, using an obscured text rewriting models. The rewriting models are guided by corresponding classifiers for the personal attributes to obscure. Adversarial training and fairness risk measurement are proposed to enhance the fairness of the generators, alleviating privacy leakage of the target attributes. The second work protects personal information in the form of open-domain textual descriptions. We further explore three feasible rewriting strategies, deleting, obscuring, and steering, for privacy-aware text rewriting. We investigate the possibility of fine-tuning a pre-trained language model for privacy-aware text rewriting. Based on our dataset, we further observe the relation of rewriting strategies to their semantic spaces in a knowledge graph. Then, a simple but effective decoding method is developed to incorporate these semantic spaces into the rewriting models. As a whole, this thesis presents a comprehensive study and the first solutions in varying settings for protecting privacy in conversations. We demonstrate that both privacy leakage detection and privacy-aware text rewriting are plausible using machine learning methods. Our contributions also include novel ideas for text alignment for natural language inference, training technologies for attribute obfuscating, and open-domain knowledge guidance to text rewriting. This thesis opens up inquiries into protecting sensitive user information in conversations from the perspective of explicit text representation.