1 paper · 1 filter
Yuzheng Hu, Fan Wu, Ruicheng Xian +5
We propose the notion of empirical privacy variance and study it in the context of differentially private fine-tuning of language models. Specifically, we show that models calibrat…