Research map / Image generation / Fast diffusion sampling
Positive-Unlabeled Diffusion Models for Preventing Sensitive Data Generation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes positive-unlabeled diffusion models that prevent sensitive data generation by approximating the normal-data ELBO from unlabeled and a small labeled sensitive set, with a non-negative risk correction and bounded BCE-style loss for sensitive data.
Problem
Diffusion models generate sensitive content because unlabeled training data contain sensitive samples; labeling all sensitive data is impractical, so only unlabeled data plus a small labeled sensitive set is available (positive-unlabeled setting).
Approach
Rewrites the normal-data distribution as unlabeled minus beta-weighted sensitive distribution, giving an unbiased PU estimate of the normal-data ELBO; maximizes that ELBO for normal data while minimizing a bounded binary-cross-entropy transform of the ELBO for sensitive data, with a non-negative correction to avoid overfitting.
Key results
- On from-scratch MNIST (even), the proposed method raises non-sensitive rate from 0.904 (unsupervised) to 0.978 while FID is 4.969 vs 3.530.
- On fine-tuned CIFAR10 (vehicles), non-sensitive rate improves from 0.741 (unsupervised) to 0.858 (proposed) with FID 18.977 vs 26.710.
- On fine-tuned Stable Diffusion 1.4 (Brad Pitt exclusion), non-sensitive rate is 0.99 for proposed vs 0.80 unsupervised and 0.84 supervised.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use PU training with unlabeled and sensitive data instead of standard unsupervised training | Diffusion models trained from scratch · MNIST (even), MNIST (odd), CIFAR10, STL10 | non-sensitive rate: 0.904$\pm$0.003 -> 0.978$\pm$0.003; 0.900$\pm$0.006 -> 0.975$\pm$0.004; 0.762$\pm$0.006 -> 0.842$\pm$0.006; 0.917$\pm$0. | Helped |
| Use PU training instead of supervised training that treats all unlabeled data as normal | Diffusion models trained from scratch · MNIST, CIFAR10, STL10 | non-sensitive rate: 0.908$\pm$0.003 -> 0.978$\pm$0.003; 0.901$\pm$0.006 -> 0.975$\pm$0.004; 0.760$\pm$0.004 -> 0.842$\pm$0.006; 0.918$\pm$0. | Helped |
| Use PU training instead of standard unsupervised training | Diffusion models trained from scratch · MNIST (even), MNIST (odd), CIFAR10 (animals), STL10 (animals) | FID: 3.530$\pm$0.137 -> 4.969$\pm$0.537; 3.379$\pm$0.034 -> 5.718$\pm$0.652; 34.448$\pm$0.504 -> 39.586$\pm$0.573; 135.999$\p | Hurt |
| Use PU training instead of supervised training that treats all unlabeled data as normal | Diffusion models trained from scratch · CIFAR10 (vehicles), STL10 (vehicles) | FID: 34.310$\pm$0.424 -> 32.463$\pm$0.970; 122.267$\pm$2.114 -> 120.633$\pm$4.256 | Helped |
| Use PU training instead of supervised training that treats all unlabeled data as normal | Diffusion models fine-tuned from pre-trained models · CIFAR10 and CelebA | non-sensitive rate: 0.763$\pm$0.006 -> 0.858$\pm$0.009; 0.892$\pm$0.005 -> 0.954$\pm$0.005; 0.868$\pm$0.020 -> 0.973$\pm$0.010; 0.907$\pm$0. | Helped |
| Use PU training instead of supervised fine-tuning | Pre-trained diffusion models · CIFAR10 (vehicles) | FID: 24.458$\pm$0.387 -> 18.977$\pm$0.381 | Helped |
| Use PU training instead of supervised fine-tuning | Pre-trained diffusion models · CIFAR10 (animals) | FID: 19.528$\pm$0.277 -> 19.415$\pm$0.915 | No clear effect |
| Use PU training instead of supervised fine-tuning | Pre-trained diffusion models · CelebA (male), CelebA (female) | FID: 43.465$\pm$2.509 -> 46.877$\pm$5.239; 30.569$\pm$1.166 -> 39.131$\pm$1.961 | Hurt |
| Increase β from the true ratio to 0.15–0.20 (ablation) | Proposed method · CIFAR10 (vehicles) | non-sensitive rate and FID | Helped |
| Set β to 0.2 instead of values at or below 0.091 (ablation) | Proposed method · CIFAR10 (animals) | non-sensitive rate | Helped |
| Use β ≤ 0.091 instead of β = 0.2 (ablation) | Proposed method · CIFAR10 (animals) | FID | Helped |
| Replace the max-based PU loss with an absolute-value-based loss (ablation) | Stable Diffusion 1.4 · Middle-aged men generation | non-sensitive rate: 0.80; 0.84; 0.99 | Helped |
Surprising findings
- Supervised training with labeled sensitive data barely improves over unsupervised training when sensitive data are present in the unlabeled set.
- The max-based non-negative correction caused divergence for Stable Diffusion, requiring the absolute-value variant.
Limitations
- Requires labeled sensitive data to represent all sensitive modes; uncovered sensitive categories (e.g., digit 9 when only 1,3,5,7 labeled) are still generated.
- Requires a hyperparameter beta (ratio of sensitive data), with performance sensitive to its value; estimation via PU methods is not explored.
- Slight FID degradation in from-scratch training, e.g., MNIST and STL10 animals.
Experimental setup
Datasets: MNIST CIFAR10 STL10 CelebA Stable Diffusion generated images (Brad Pitt / middle-aged men)
Benchmarks: MNIST CIFAR10 STL10 CelebA
Metrics: non-sensitive rate FID
Baselines: unsupervised diffusion training supervised diffusion training with sensitive data
Builds on: denoising diffusion probabilistic models unbiased PU learning non-negative PU risk estimator classifier-free guidance
Closest papers by meaning
- Asymptotically Exact Negative Guidance of Diffusion Models via Positive-Unlabeled Learning (NeurIPS 2026 · similarity 0.89)
- Training-Free Safe Denoisers for Safe Use of Diffusion Models (NeurIPS 2025 · similarity 0.89)
- Erasing Undesirable Influence in Diffusion Models (CVPR 2025 · similarity 0.88)
- Diffusion-NPO: Negative Preference Optimization for Better Preference Aligned Generation of Diffusion Models (ICLR 2025 · similarity 0.88)
- Towards Resilient Safety-driven Unlearning for Diffusion Models against Downstream Fine-tuning (NeurIPS 2025 · similarity 0.88)
- The Illusion of Unlearning: The Unstable Nature of Machine Unlearning in Text-to-Image Diffusion Models (CVPR 2025 · 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.