Re-examining low-rank adaptation for private LLM fine-tuning
Organizations: Dalhousie University & Vector Institute
Abstract
Privacy is a central concern when fine-tuning large language models (LLMs) on sensitive data, and differentially private stochastic gradient descent (DP-SGD) provides formal privacy guarantees. When performance at a fixed privacy level largely saturates on established benchmarks such as GLUE, we ask whether the same performance can be reached with less computation. Both theory and practice show that lower-rank models are better suited to DP training, a property especially relevant for LLMs, whose fine-tuning gradients exhibit a strong low-rank structure. Methods such as DP-LoRA exploit this by expressing each layer's weight update as a sum of only a few non-zero rank-one components. However, we argue that while having few non-zero components is important, the isotropic noise injected by DP-SGD inflates the singular values of the gradient matrix, disrupting their naturally fast decay. We investigate whether this noise-induced spectral distortion slows optimization. Experiments on language classification (GLUE with RoBERTa) and table-to-text generation (E2E and DART with Qwen and Llama models up to 4B parameters) show that partially restoring the original singular-value profile significantly accelerates DP-SGD convergence without compromising privacy guarantees. Our code is available at https://github.com/alidadsetan/re-exam-private-low-rank.
Figures & tables
| Task | Model | Results at Step 200 | Results at Step 400 | ||||||
|---|---|---|---|---|---|---|---|---|---|
| DP-LoRA Acc. | Ours Acc. | Speedup | DP-LoRA Acc. | Ours Acc. | Speedup | ||||
| SST-2 | Base | 90.8 0.3 | 27.3% | 0.57 | 92.1 0.3 | 17.5% | 0.55 | ||
| Large | 93.1 0.2 | 62.5% | 1.47 | 94.2 0.2 | 15.6% | 0.49 | |||
| MNLI | Base | 71.2 3.5 | 46.4% | 0.55 | 79.9 0.5 | 34.5% | 0.56 | ||
| Large | 78.8 8.3 | 15.6% | 0.17 | 85.9 0.1 | 19.3% | 0.30 | |||
| QNLI | Base | 78.2 1.3 | 75.0% | 1.64 | 84.4 0.3 | 40.5% | 1.17 | ||
| Clipping method | Step 50 | Step 100 | Step 150 | |||
|---|---|---|---|---|---|---|
| DP-LoRA | Ours | DP-LoRA | Ours | DP-LoRA | Ours | |
| Fixed | 50.92 | 50.92 | 50.92 | 89.45 | 89.89 | 90.30 |
| Fixed | 50.92 | 50.92 | 50.92 | 89.54 | 89.54 | 90.73 |
| Fixed | 50.92 | 50.92 | 50.92 | 80.96 | 50.92 | 90.14 |
| AUTO-S ( ) | 50.92 | 50.92 | 50.92 | 89.36 | 89.29 | 90.64 |
| Task | Model | Budget (min) | DP-LoRA | DOPPLER | DiSK | Ours | ||||
|---|---|---|---|---|---|---|---|---|---|---|
| Steps | Acc. (%) | Steps | Acc. (%) | Steps | Acc. (%) | Steps | Acc. (%) | |||
| Time budget corresponding to 200 DP-LoRA steps | ||||||||||
| MNLI | Base | 26.79 | 200 | 200 | 97 | 197 | ||||
| Large | 77.34 | 200 | 198 | 96 | 197 | |||||
| SST-2 | Base | 26.74 | 200 | 199 | 97 | 197 | ||||
| Large | 76.98 | 200 | 198 | 96 | 196 | |||||
| Task | Model | BLEU | ROUGE-L | METEOR | NIST | CIDEr | |||||
|---|---|---|---|---|---|---|---|---|---|---|---|
| DP-LoRA | Ours | DP-LoRA | Ours | DP-LoRA | Ours | DP-LoRA | Ours | DP-LoRA | Ours | ||
| E2E | Qwen 0.6B | 23.17 | 37.35 | 46.78 | 53.52 | 0.509 | 0.628 | 2.34 | 4.60 | 0.67 | 1.11 |
| Qwen 1.7B | 36.01 | 36.95 | 53.57 | 54.42 | 0.609 | 0.624 | 4.64 | 4.71 | 1.08 | 1.14 | |
| Qwen 4B | 40.00 | 40.88 | 55.19 | 55.74 | 0.659 | 0.663 | 4.83 | 5.03 | 1.27 | 1.36 | |
| Llama 1B | 22.15 | 36.33 | 44.24 | 52.86 | 0.476 | 0.626 | 2.59 | 4.62 | 0.47 | 1.13 | |
| Llama 3B | 10.96 | 26.33 | 25.22 | 44.36 | 0.268 | 0.516 | 0.93 | 3.48 | 0.22 | 0.64 | |
Appendix figures & tables8 assets
Supplementary material from the paper’s appendix.
Appendix
| Model | Method | Train Time | Overhead |
|---|---|---|---|
| RoBERTa Base | Regular DP-SGD | 1846 | – |
| Naïve denoising implementation | 1899 | 2.87% | |
| Efficient denoising implementation | 1861 | 0.81% | |
| RoBERTa Large | Regular DP-SGD | 5658 | – |
| Naïve denoising implementation | 5781 | 2.17% | |
| Efficient denoising implementation | 5700 | 0.74% |
| Dataset | Target | Step 50 | Step 100 | Step 150 | Step 200 | Step 250 | Step 300 | Step 350 | Step 400 |
|---|---|---|---|---|---|---|---|---|---|
| SST-2 | Clipped | 9.38% | 9.91% | 5.08% | 1.33% | 1.17% | 0.91% | 1.15% | 0.69% |
| Unclipped | 6.31% | 4.66% | 3.54% | 1.10% | 0.94% | 0.77% | 0.98% | 0.57% | |
| MNLI | Clipped | 16.74% | 16.71% | 16.98% | 16.83% | 9.49% | 8.35% | 7.77% | 6.03% |
| Unclipped | 16.31% | 16.28% | 16.29% | 12.03% | 6.87% | 6.11% | 5.50% | 4.13% |
Privacy-Preserving Language Models
Privacy-preserving language models protect sensitive data during training, inference, personalization, and information sharing.
See all 191 papers