Position: Generative AI and Differential Privacy --- A Perfect Match

Jonas Kuntzer, Johannes Kaiser, Tamara T. Müller, Anneliese Riess, Kristian Schwethelm, Daniel Rueckert

TL;DR: The paper highlights DP's potential for addressing several critical GenAI issues and points out the lack of research on this.

Abstract: Generative AI (GenAI) has evolved into a transformative technology whose unprecedented growth and public exposure have revealed challenging issues ranging from privacy protection to reducing factual inaccuracies and hallucinations, model security risks, legal complications, and a lack of interpretability. This $\textit{position paper}$ examines how Differential Privacy (DP), a mathematical privacy protection framework, can address privacy concerns $\textit{and}$ other systemic challenges beyond privacy in GenAI. We argue that DP is a versatile and underutilized tool with significant potential to address many critical GenAI issues. To argue our claim, we connect the core principle of DP to these issues, evaluate existing research, and pose relevant research questions.

Primary Area: Research Priorities, Methodology, and Evaluation

Keywords: generative AI, differential privacy, privacy protection, factual inaccuracies, hallucinations, security, backdoors, "right to erasure", interpretability, memorization, data pricing, right to erasure, right to be forgotten, unlearning

Submission Number: 178