Research map / Federated learning and privacy / Differential privacy
Differentially Private Fine-Tuning of Diffusion Models
Yu-Lin Tsai, Yizhe Li, Chia-Mu Yu, Xuebin Ren, Po-Yu Chen, Zekai Chen, Francois Buet-Golfouse
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 LoRA to QKV attention matrices (ablation) | DP-LoRA · CelebA-32 | FID: 11.78 -> 7.71 | Improved |
| add LoRA to the output projection layer (ablation) | DP-LoRA · CelebA-32 | FID: 9.82 -> 7.71 | Improved |
| add LoRA to the output projection layer (ablation) | DP-LoRA · CelebA-64 | FID: 11.24 -> 8.41 | Improved |
| increase noise multiplicity from k=1 to k=4 (ablation) | DP-LoRA · CelebA-32 | FID: 13.75 -> 7.71 | Improved |
| increase noise multiplicity from k=4 to k=8 (ablation) | DP-LoRA · CelebA-64 | FID: 8.41 -> 8.94 | Hurt |
| increase noise multiplicity from k=4 to k=8 (ablation) | DP-LoRA · CelebA-32 | FID: 7.71 -> 7.32 | Improved |
| raise LoRA rank from 16 to 64 (ablation) | DP-LoRA · CelebA-32 | FID: 7.71 -> 10.16 | Hurt |
| lower LoRA rank from 16 to 8 (ablation) | DP-LoRA · CelebA-32 | FID: 7.71 -> 8.09 | No clear effect |
| use parameter-efficient LoRA fine-tuning instead of DP-LDM full fine-tuning | DP-LDM · CelebA-64 eps=10 | FID: 14.3 -> 8.4 | Improved |
| use LoRA latent diffusion instead of DP-LDM | DP-LDM · CelebA-HQ eps=1 | FID: 25.6 -> 20.0 | Improved |
| use LoRA fine-tuning instead of DP-LDM | DP-LDM · CIFAR-10 ResNet9 eps=10 | accuracy: 65.3 -> 73.98 | Improved |
| use LoRA latent diffusion instead of vanilla pixel-space DPDM | DPDM · Fashion-MNIST eps=10 | accuracy: 86.2 -> 72.7 | Hurt |
Closest papers by meaning
- RAPID: Retrieval Augmented Training of Differentially Private Diffusion Models (ICLR 2025 · similarity 0.91)
- In-distribution Public Data Synthesis with Diffusion Models for Differentially Private Image Classification (CVPR 2024 · similarity 0.91)
- Optimal Differentially Private Model Training with Public Data (ICML 2024 · similarity 0.90)
- Differentially Private Prototypes for Imbalanced Transfer Learning (AAAI 2025 · similarity 0.89)
- DiSK: Differentially Private Optimizer with Simplified Kalman Filter for Noise Reduction (ICLR 2025 · similarity 0.89)
- Privacy without Noisy Gradients: Slicing Mechanism for Generative Model Training (NeurIPS 2024 · similarity 0.89)
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.
