MGE: A Membership Guided Data Extraction Attack against Diffusion Models
Organizations: Renmin University of China, Beijing, China
Abstract
Recently, diffusion models have achieved remarkable performance in image generation. However, their inherent memory properties introduce a potential safety risk of training data leakage, i.e., data extraction attack. To illustrate such risk, we propose a membership guided extraction (MGE) framework to effectively extract training data from a pre-trained diffusion model. Our core idea is to formulate the data extraction as a conditional generation, where the condition signal is from a membership inference classifier. To improve the success rate and efficiency of MGE, we propose adversarial gradient and parallel branch decoding. Our experiments on image generation benchmarks demonstrate that our MGE can extract hundreds of images, which shows that more attention should be paid to privacy-preserving of diffusion model.
Figures & tables
| ADM | |||||||||
|---|---|---|---|---|---|---|---|---|---|
| CIFAR10 | CIFAR100 | Tiny-IN | |||||||
| Method | Quantity | ASR | Efficiency | Quantity | ASR | Efficiency | Quantity | ASR | Efficiency |
| MGE | / | / | / | ||||||
| DDPM ( Carlini et al. 2023 ) | 55 / | 0.042% | 0.08 | 129 / | 0.098% | 0.18 | 41 / | 0.031% | 0.06 |
| DDIM ( Carlini et al. 2023 ) | 26 / | 0.020% | 0.37 | 52 / | 0.040% | 0.57 | 24 / | 0.018% | 0.25 |
| EDM | |||||||||
| MIA Classifier | Quantity | ASR | Efficiency | AUC | TPR@1%FPR |
|---|---|---|---|---|---|
| MGE-NA | 33 / | 0.025% | 0.22 | 84.7 | 6.85 |
| MGE-SecMI | 70 / | 0.053% | 0.03 | 88.1 | 9.11 |
| MGE-PIA | 141 / | 0.107% | 0.78 | 91.9 | 24.7 |
| Augmentation Strength | Quantity | ASR | Efficiency |
|---|---|---|---|
| without augmentation | 18 / | 0.014% | 0.2 |
| 1 | 58 / | 0.044% | |
| 2 | 57 / | 0.043% | 0.48 |
| 4 | / | 0.78 |
| Branch Nums | 1 | 2 | 4 | 8 |
|---|---|---|---|---|
| Quantity | 73 / | 60 / | / | 65 / |
| ASR | 0.056% | 0.046% | 0.050% | |
| Efficiency | 0.35 | 0.36 | 0.50 |
| Approximation Strategy | Quantity | ASR | Efficiency | Duplicates |
|---|---|---|---|---|
| Ours | 0.107% | 0.783 | 1094 | |
| 1/5 total steps | 0.011% | 0.020 | 13859 | |
| 1/2 total steps | 0.012% | 0.011 | 13844 | |
| Full steps | 0.014% | 0.007 | 13885 |
Appendix figures & tables10 assets
Supplementary material from the paper’s appendix.
Appendix
| Guidance Scale | 0.0 | 10.0 | 20.0 | 40.0 | 50.0 | 60.0 | 70.0 | 80.0 | 90.0 |
|---|---|---|---|---|---|---|---|---|---|
| Quantity | 6 / | 14 / | 39 / | 72 / | 71 / | 70 / | / | 58 / | 56 / |
| ASR | 0.005% | 0.011% | 0.030% | 0.055% | 0.054% | 0.053% | 0.044% | 0.043% | |
| Efficiency | 0.03 | 0.08 | 0.22 | 0.40 | 0.39 | 0.39 | 0.32 | 0.31 |
| 1.0 | 2.0 | 5.0 | 10.0 | 15.0 | |
|---|---|---|---|---|---|
| Quantity | |||||
| ASR (%) | 0.010 | 0.010 | 0.020 | 0.041 | 0.107 |
| Efficiency | 0.06 | 0.06 | 0.15 | 0.31 | 0.78 |
| AUC | 91.7 | 91.8 | 91.9 | 91.9 | 91.9 |
| TPR@1%FPR | 24.6 | 24.7 | 24.7 | 24.7 | 24.7 |
| Guidance Phase | Quantity | ASR | Efficiency |
|---|---|---|---|
| / | |||
| 42 / | 0.032% | 0.42 | |
| 7 / | 0.005% | 0.03 | |
| 47 / | 0.036% | 0.20 |
| Setting | Quantity | ASR | Efficiency | Duplicates |
|---|---|---|---|---|
| Non-deduplicated | 0.159% | 1.18 | 775 | |
| Deduplicated | 0.171% | 0.20 | 548 |
| Method | Quantity | ASR | Efficiency |
|---|---|---|---|
| SIDE | 96 | 0.073% | 0.20 |
| MGE | 141 | 0.107% | 0.78 |
| Guidance gradient | Quantity | ASR | Efficiency |
|---|---|---|---|
| Zeroth-order | 118 | 0.090% | 0.44 |
| Backpropagation | 141 | 0.107% | 0.78 |
| Setting | Quantity | ASR | Efficiency |
|---|---|---|---|
| Full MGE | 27 | 0.0108% | 0.075 |
| Multistep (1/5 steps) | 5 | 0.0020% | 0.003 |
| Without branching | 14 | 0.0056% | 0.032 |
| Run | Unique extractions |
|---|---|
| Seed 3 | 358 |
| Seed 36 | 346 |
| Seed 42 | 351 |
| Mean sample std. |