Reproduce “LLM-PBE: Assessing Data Privacy in Large Language Models” on OpenResearch

About this paper

Qinbin Li, Junyuan Hong, Chulin Xie, Jeffrey Tan, Rachel Xin, Junyi Hou, Xavier Yin, Zhun Wang and 5 more

Researchers from UC Berkeley, UT Austin, UIUC, NUS, Center for AI Safety, and University of Chicago introduce LLM-PBE, a comprehensive toolkit and framework for systematically assessing data privacy risks in Large Language Models. This work provides empirical insights into how factors like model size, data characteristics, and various attack vectors influence leakage, and evaluates the efficacy of privacy-enhancing technologies.

Key insights

• Increasing LLM size and pretraining data volume leads to significantly higher data memorization and privacy leakage, with privacy risk escalation often outpacing utility gains.
• Specific data characteristics, such as the position of private data within a sample (end-of-sequence) and certain data types (e.g., locations), exhibit heightened vulnerability to leakage.
• Differential Privacy (DP-SGD) offers a more balanced trade-off between privacy protection and model utility compared to scrubbing, while simple defensive prompts showed limited effectiveness against prompt leakage.

Summary by alphaXiv.