@inproceedings{yu2022differentially, author = {Yu, Da and Naik, Saurabh and Backurs, Arturs and Gopi, Sivakanth and Inan, Huseyin and Kamath, Gautam and Kulkarni, Janardhan (Jana) and Lee, Yin Tat and Manoel, Andre and Wutschitz, Lukas and Yekhanin, Sergey and Zhang, Huishuai}, title = {Differentially private fine-tuning of language models}, booktitle = {ICLR 2022}, year = {2022}, month = {April}, abstract = {We give simpler, sparser, and faster algorithms for differentially private fine-tuning of large-scale pre-trained language models, which achieve the state-of-the-art privacy versus utility tradeoffs on many standard NLP tasks. We propose a meta-framework for this problem, inspired by the recent success of highly parameter-efficient methods for fine-tuning. Our experiments show that differentially private adaptations of these approaches outperform previous private algorithms in three important dimensions: utility, privacy, and the computational and memory cost of private training. On many commonly studied datasets, the utility of private models approaches that of non-private models. For example, on the MNLI dataset we achieve an accuracy of 87.8% using RoBERTa-Large and 83.5% using RoBERTa-Base with a privacy budget of ϵ=6.7. In comparison, absent privacy constraints, RoBERTa-Large achieves an accuracy of 90.2%. Our findings are similar for natural language generation tasks. Privately fine-tuning with DART, GPT-2-Small, GPT-2-Medium, GPT-2-Large, and GPT-2-XL achieve BLEU scores of 38.5, 42.0, 43.1, and 43.8 respectively (privacy budget of ϵ=6.8,δ= 1e-5) whereas the non-private baseline is 48.1. All our experiments suggest that larger models are better suited for private fine-tuning: while they are well known to achieve superior accuracy non-privately, we find that they also better maintain their accuracy when privacy is introduced.}, url = {http://approjects.co.za/?big=en-us/research/publication/differentially-private-fine-tuning-of-language-models/}, }