“What did you say?”: Extracting unintentional secrets from predictive text learning systems

摘要

As a primary form of communication, text is used widely for online communications, including e-mail conversations, mobile text messaging, chatroom and forum discussions. Modern systems include facilities such as predictive text, recently implemented using deep learning algorithms, to estimate the next word to be written based on previous historical entries. However, we often enter sensitive information such as passwords using the same input devices - namely, smartphone soft keyboards. In this paper, we explore the problem of deep learning models which memorise sensitive training data, and how secrets can be extracted from predictive text models. We propose a general black-box attack algorithm to accomplish this for all kinds of memorised sequences, discuss mitigations and countermeasures, and explore how this attack vector could be deployed on an Android or iOS mobile device platforms as part of target reconnaissance.