Research map / Federated learning and privacy / Personalized and heterogeneous FL
Rethinking LoRA for Privacy-Preserving Federated Learning in Large Models
Jin Liu, Yinbin Miao, Ning Xi, Junkang Liu
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 |
|---|---|---|---|
| add local alternating updates of A and B (ablation) | DP-LoRA · Tiny-ImageNet (Swin-B, eps=3) | test accuracy (%): 30.64 -> 53.07 | Improved |
| add local alternating updates of A and B (ablation) | DP-LoRA · QNLI (RoBERTa-Base, eps=3) | test accuracy (%): 86.31 -> 88.92 | Improved |
| add Gaussian low-pass filter (ablation) | LA-LoRA without filter · Tiny-ImageNet (Swin-B, eps=3) | test accuracy (%): 53.07 -> 61.97 | Improved |
| add Gaussian low-pass filter (ablation) | DP-LoRA · Tiny-ImageNet (Swin-B, eps=3) | test accuracy (%): 30.64 -> 49.85 | Improved |
| add Gaussian low-pass filter (ablation) | DP-LoRA · SST-2 (RoBERTa-Base, eps=3) | test accuracy (%): 92.36 -> 92.52 | No clear effect |
| replace Doppler filter with Gaussian filter (ablation) | LA-LoRA · Tiny-ImageNet (Swin-B, eps=3) | test accuracy (%): +4.61% | Improved |
| vary local alternating schedule (1,2,5 steps per factor, A-first or B-first) (ablation) | LA-LoRA · CIFAR-100/Tiny-ImageNet (Swin-B) | test accuracy (%): best-worst gap below 0.7% | No clear effect |
| increase LoRA rank 8 -> 32 (ablation) | LA-LoRA · CIFAR-100 (Swin-B, eps=1) | test accuracy (%): 72.88 -> 75.34 | Improved |
| use LA-LoRA instead of RoLoRA | RoLoRA · Tiny-ImageNet (Swin-B, eps=1) | test accuracy (%): 43.85 -> 60.68 | Improved |
| use LA-LoRA instead of RoLoRA | RoLoRA · MNLI (RoBERTa-Base, eps=3) | test accuracy (%): 82.12 -> 82.99 | Improved |
| use LA-LoRA on Llama-2-7B instead of RoLoRA | RoLoRA · GLUE avg (Llama-2-7B, eps=1) | average accuracy (%): 88.66 -> 89.36 | Improved |
| use local alternating updates | DP-LoRA · Max Hessian eigenvalue (Swin-B CIFAR-100, eps=1) | max Hessian eigenvalue: 101.62 -> 64.77 | Improved |
Closest papers by meaning
- Improving LoRA in Privacy-preserving Federated Learning (ICLR 2024 · similarity 0.95)
- Differentially Private Federated Low Rank Adaptation Beyond Fixed-Matrix (NeurIPS 2025 · similarity 0.93)
- FedSVD: Adaptive Orthogonalization for Private Federated Learning with LoRA (NeurIPS 2025 · similarity 0.92)
- Privacy-Preserving Low-Rank Adaptation Against Membership Inference Attacks for Latent Diffusion Models (AAAI 2025 · similarity 0.90)
- Fed-SB: A Silver Bullet for Extreme Communication Efficiency and Performance in (Private) Federated LoRA Fine-Tuning (NeurIPS 2026 · similarity 0.88)
- Adaptive LoRA Experts Allocation and Selection for Federated Fine-Tuning (NeurIPS 2025 · similarity 0.88)
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.
