Core Concepts
Differential privacy can be effectively used to protect tabular data in In-Context Learning, ensuring privacy while maintaining performance.
Abstract
In the study, the authors investigate the use of differential privacy (DP) to safeguard tabular data in In-Context Learning (ICL). They propose two frameworks, Local Differentially Private Tabular-based ICL (LDP-TabICL) and Global Differentially Private Tabular-based ICL (GDP-TabICL), to generate demonstration examples with provable privacy guarantees. The experiments show that DP-based ICL can maintain data privacy while achieving comparable performance to non-DP models across various settings. The study focuses on protecting sensitive information contained in tabular datasets used for ICL tasks by leveraging DP mechanisms. By using noise injection and aggregation techniques, the authors demonstrate how LDP and GDP can ensure data privacy without compromising model performance.
Stats
LLM: Llama-2-13B
Datasets: adult, bank, blood, calhousing, car, diabetes, heart, jungle
Privacy Budgets: ϵ = {1, 5, 10, 25, 50}
Key Metrics: Accuracy scores for LDP-TabICL and baseline models
Quotes
"Understanding how to protect the underlying tabular data used in ICL is a critical area of research."
"DP-based ICL can protect the privacy of the underlying tabular data while achieving comparable performance to non-LLM baselines."