Prompt Privacy in the era of Large Language Models

In recent years, large language models (LLMs) have significantly impacted text-related tasks, including
translation, code generation, and answering questions. However, due to their computational demands, these
models are typically available only through closed doors-APIs, with opaque operations that obscure data usage.
This creates privacy concerns for users who frequently engage with these models, as their interactions may
inadvertently reveal sensitive personal information. These exchanges are expressed through prompts, which,
despite containing confidential data, provide essential context for the models to perform tasks. Our research
concentrates on examining privacy issues in user dialogues with proprietary LLMs like ChatGPT. We plan to
process and study extensive datasets of user-model interactions to understand the extent of privacy breaches.
Additionally, we aim to annotate these conversations by sensitive areas such as finance or healthcare to
enhance our analysis.

Faculty Supervisor:

Golnoosh Farnadi

Student:

Partner:

Mouvement des caisses Desjardins

Discipline:

Computer science

Sector:

Finance and Insurance

University:

McGill University

Program: