AI Paper Atlas

Research map / Federated learning and privacy / Differential privacy

Differentially Private Fine-tuning of Language Models

Da Yu, Saurabh Naik, Arturs Backurs, Sivakanth Gopi, Huseyin A Inan, Gautam Kamath, Janardhan Kulkarni, Yin Tat Lee, Andre Manoel, Lukas Wutschitz, Sergey Yekhanin, Huishuai Zhang

ICLR 2022 · Federated learning and privacy › Differential privacy · approach: Differentially private synthetic data · 46 citations (OpenAlex)

Read the paper

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
replace full DPSGD fine-tuning with parameter-efficient private fine-tuningRoBERTa-Large · MNLI, SST-2, QQP, QNLIaccuracy: 86.1, 93.0, 86.7, 90.0 -> 87.8, 95.3, 87.4, 90.8Improved
freeze pre-trained weights and train LoRA parameters rather than use RGP decomposition and projectionprivate fine-tuning of RoBERTa-Large · SST-2memory: 9.1 -> 6.1 GBImproved
freeze pre-trained weights and train LoRA parameters rather than use RGP decomposition and projectionprivate fine-tuning of RoBERTa-Large · SST-2seconds per epoch: 296 -> 271Improved
increase backbone size from RoBERTa-Base to RoBERTa-LargeDP LoRA fine-tuning · MNLIaccuracy: 83.5 -> 87.8Improved
increase model size from GPT-2-Medium to GPT-2-XLDP LoRA fine-tuning · DARTBLEU: 42.0 -> 43.8Improved
increase model size from GPT-2-Large to GPT-2-XLDP LoRA fine-tuning · E2EBLEU: 66.7 -> 66.1Hurt
use Adapter instead of LoRADP fine-tuning of RoBERTa-Base · GLUE average accuracyaccuracy: 87.2 -> 87.3Improved
use LoRA instead of AdapterDP fine-tuning of RoBERTa-Large · GLUE average accuracyaccuracy: 89.7 -> 90.3Improved
use full precision and larger batches for larger datasetsDP fine-tuning of RoBERTa-Large · MNLIaccuracy: 87.8 -> 89.0Improved
use full precision and larger batches for larger datasetsDP fine-tuning of RoBERTa-Base · GLUE average accuracyaccuracy: 87.2 -> 88.2Improved
use a larger privacy budget, from ε=1 to ε=3DP LoRA with RoBERTa-Large · MNLIaccuracy: 86.8 -> 87.4Improved
use a larger privacy budget, from ε=1 to ε=3DP LoRA with RoBERTa-Large · SST-2accuracy: 93.1 -> 94.1Improved

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy