<ul><li>Differentially private (DP) language model inference is utilized for creating private synthetic text using large language models (LLM).</li><li>Clustering input data before selecting inference batches improves the quality of privately generated text, especially for heterogeneous topics.</li><li>A new algorithm aggregates next token statistics by privately computing medians instead of averages, benefiting from decreased local sensitivity.</li><li>This approach offers high-quality synthetic data with lower privacy cost compared to the previous state-of-the-art method, showcasing improvements in representativeness metrics and downstream task performance.</li></ul>

Clustering and Median Aggregation Improve Differentially Private Inference

Discover more