Reproduce “SoK: The Privacy Paradox of Large Language Models: Advancements, Privacy Risks, and Mitigation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yashothara Shanmugarasa, Ming Ding, M.A.P Chamikara, Thierry Rakotoarivelo
Researchers from CSIRO's Data61 present a Systematization of Knowledge (SoK) that comprehensively categorizes privacy risks in Large Language Models (LLMs) across training data, user prompts, generated outputs, and autonomous agents. The work details various attacks and mitigation strategies, emphasizing the underexplored challenges related to user interaction and agent deployment.
Key insights
- LLM privacy risks extend significantly beyond training data concerns, encompassing the entire lifecycle of LLM engagement, from user prompts to autonomous agent interactions.
- LLMs' advanced capabilities allow for sophisticated inference of sensitive information from seemingly innocuous inputs, creating new vectors for privacy breaches.
- Mitigation strategies for privacy challenges related to user prompts, LLM-generated outputs, and autonomous agents are significantly less mature compared to those addressing training data privacy.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.