Large Language Models for Managing Online Fingerprint

Abstract

Today, many are unaware of how much of their personal information is publicly available on the web, which has become an increasingly important issue among internet users. This thesis builds on the work of the preceding Capstone project and uses the open-source Online Privacy Pilot tool as a case study to explore how large language models can be incorporated into the tool to enhance its functionality and assist users in managing their online fingerprint.

Based on our evaluation of the Mistral and Llama 2 models, we selected Mistral and incorporated it into three features of the Online Privacy Pilot tool: generating recommended positive filters, clustering user profile entries, and creating informative snippets for these entries. The recommended positive filters are generated based on the entries in the user profile and allow the user to provide relevance feedback to the tool if they choose to add them to the search query. Additionally, we selected and proposed a total of 13 cluster labels for use in the tool's clustering feature. To address ethical and legal considerations, especially concerning user intent and data privacy, we implemented an additional step when adding footprints to the user profile, guiding users to store only personally relevant footprints.