Research map / Federated learning and privacy / Differential privacy
Sparsity-Preserving Differentially Private Training of Large Embedding Models
Badih Ghazi, Yangsibo Huang, Pritish Kamath, Ravi Kumar, Pasin Manurangsi, Amer Sinha, Chiyuan Zhang
Read the paperCheck your idea against 137,217 papers
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 changed | Where | Measured | Effect |
|---|---|---|---|
| use DP-AdaFEST instead of LoRA for embedding gradient size | RoBERTa word embeddings · SST-2 eps=1.0 | gradient size reduction: 5.91 -> 14.59 at utility loss 0.001 | Improved |
| use DP-AdaFEST instead of LoRA at utility loss 0.01 | RoBERTa word embeddings · SST-2 eps=1.0 | gradient size reduction: 47.28 -> 53.90 | Improved |
| larger vocabulary model (XLM-R 250k vs RoBERTa 50k) | DP-AdaFEST · XNLI vs SST-2 eps=1.0 | gradient size reduction: 53.90 -> 152.94 at utility loss 0.01 | Improved |
| combine DP-FEST preselection with DP-AdaFEST (DP-AdaFEST+) | DP-AdaFEST · Criteo-Kaggle | gradient size reduction: >10^6 | Improved |
| combine DP-FEST with DP-AdaFEST | DP-AdaFEST · Criteo-time-series | utility/efficiency trade-off | Improved |
| apply exponential selection sparsity to DP-SGD | DP-SGD · Criteo-Kaggle, SST-2, QNLI, QQP | utility | Hurt |
| DP-AdaFEST instead of DP-FEST | DP-FEST · Criteo-Kaggle | gradient size reduction at same utility | Improved |
| DP-AdaFEST instead of DP-FEST on time-series data | DP-FEST · Criteo-time-series eps=1.0 | gradient reduction at same utility | Improved |
| DP-FEST frequency from streaming sums instead of first day only (ablation) | DP-FEST · Criteo-time-series | utility | Improved |
| streaming-based instead of all-day frequency for DP-FEST (ablation) | DP-FEST · Criteo-time-series | utility | No clear effect |
| increase noise ratio sigma1/sigma2 (ablation) | DP-AdaFEST · Criteo-Kaggle and SST-2 | accuracy: best utility at ratio 5 or 10 | Improved |
| increase noise ratio sigma1/sigma2 (ablation) | DP-AdaFEST · Criteo-Kaggle and SST-2 | gradient density | Hurt |
Closest papers by meaning
- Do not Let Privacy Overbill Utility: Gradient Embedding Perturbation for Private Learning (ICLR 2021 · similarity 0.89)
- Differentially Private Optimization with Sparse Gradients (NeurIPS 2024 · similarity 0.89)
- Differentially Private Stochastic Gradient Descent with Fixed-Size Minibatches: Tighter RDP Guarantees with or without Replacement (NeurIPS 2024 · similarity 0.88)
- Differentially private optimization for non-decomposable objective functions (ICLR 2025 · similarity 0.88)
- Certification for Differentially Private Prediction in Gradient-Based Training (ICML 2025 · similarity 0.88)
- Nearly Tight Black-Box Auditing of Differentially Private Machine Learning (NeurIPS 2024 · similarity 0.87)
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.
