Abstract
dc:description.abstractLarge multimodal language models have been shown to memorize and leak pre-training data, raising serious user privacy and information security concerns. While data leaks should be prevented, it is also crucial to examine the trade-off between the privacy protection and model utility of proposed approaches. In this thesis, we introduce PrivQA— a multimodal benchmark to assess the capabilities of language models to follow privacy protection instructions such that a selected group of personal information is protected while remaining useful to respond to other queries. PrivQA consists of both annotated human-AI dialogues as well as repurposed queries from existing question-answering (QA) datasets. We propose an access control instruction technique to iteratively self-moderate language models’ responses, which significantly improves privacy. However, through a series of red-teaming experiments, we find that adversaries can easily circumvent these protections with simple jailbreaking methods through textual and/or image inputs. We believe PRIVQA has the potential to support the development of new models with improved privacy protections, as well as the adversarial robustness of these protections.
Degree
thesis:*- Level thesis:degree_level
- Masters
- Department dc:contributor.department
- Interactive Computing
- Grantor dc:publisher
- Georgia Institute of Technology
- Year dc:date.issued
- 2024
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Mendes, Ethan
- Advisor dc:contributor.advisor
-
- Ritter, Alan
- Committee members dc:contributor.committeemember
-
- Das, Sauvik
- Xu, Wei
Subjects
dc:subject × 5Rights
- Language dc:language.iso
- en_US
Identifiers
dc:identifier.*- Handle dc:identifier.uri
- https://hdl.handle.net/1853/75327