Research map / Label-efficient and robust learning / OOD and anomaly detection
Leveraging Contaminated Datasets to Learn Clean-Data Distribution with Purified Generative Adversarial Networks
Bowen Tian, Qinliang Su, Jianxing Yu
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 a discriminator that distinguishes target from contamination and train purification jointly with generation | PuriGAN · MNIST, F-MNIST, SVHN, CelebA, CIFAR-10 (γp=0.4, γc=0.2) | FID: MNIST: 10.35 ± 3.1 -> 9.51 ± 1.0; F-MNIST: 44.43 ± 5.4 -> 34.86 ± 3.2; SVHN: 21.08 ± 2.5 -> 19.63 ± 3.6; CelebA: 42.37 ± | Improved |
| Use a three-level instead of two-level discriminator (ablation) | PuriGAN · MNIST, F-MNIST, SVHN, CelebA, CIFAR-10 (γp=0.4, γc=0.2) | FID: MNIST: 9.71 ± 1.9 -> 9.51 ± 1.0; F-MNIST: 37.61 ± 4.6 -> 34.86 ± 3.2; SVHN: 20.32 ± 2.4 -> 19.63 ± 3.6; CelebA: 36.43 ± | Improved |
| Set the two-level discriminator weight λ to 3 or 5 rather than 1 (ablation) | PuriGAN₁ · MNIST, F-MNIST, SVHN at γp=0.1, 0.3, 0.5 | FID: MNIST γp=0.1: 9.04 ± 1.5 -> 9.16 ± 1.7 (λ=3), 9.41 ± 1.6 (λ=5); SVHN γp=0.1: 17.79 ± 2.1 -> 17.41 ± 2.0 (λ=3), 16.90 ± 1 | Mixed |
| Set generator target c to 0.6 or 0.7 rather than 0.5 (ablation) | PuriGAN · MNIST, F-MNIST, SVHN at γp=0.1, 0.3, 0.5 | FID: MNIST γp=0.1: 9.04 ± 1.5 -> 9.23 ± 1.5 (c=0.6), 9.34 ± 1.5 (c=0.7); F-MNIST γp=0.3: 37.17 ± 3.2 -> 36.85 ± 3.1 (c=0.6), | Mixed |
| Increase the available contamination examples | PuriGAN · Image generation across considered datasets | FID | Improved |
| Increase the contamination ratio in training | LSGAN and NDA · MNIST, F-MNIST, SVHN, CIFAR-10 | FID | Hurt |
| Estimate π with a small error rather than using its exact value (ablation) | PuriGAN · Image generation across evaluated scenarios | FID | Hurt |
| Use PuriGAN instead of PAN for classification | PU-learning · 20News, IMDB, MNIST, CIFAR-10 | F1: 20News: 81.1 -> 85.7; IMDB: 77.1 -> 79.7; MNIST: 96.5 -> 96.8; CIFAR-10: 87.2 -> 88.7 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- PU-NDA — PUnDA: Probabilistic Unsupervised Domain Adaptation for Knowledge Transfer Across Visual Categories (ICCV 2017) · better on 2 of 2 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Contamination-Resilient Anomaly Detection via Adversarial Learning on Partially-Observed Normal and Anomalous Data (ICML 2024 · similarity 0.88)
- On Positive-Unlabeled Classification From Corrupted Data in GANs (TPAMI 2025 · similarity 0.87)
- Diffusion Models for Adversarial Purification (ICML 2022 · similarity 0.87)
- Adversarial Purification by Consistency-Aware Latent Space Optimization on Data Manifolds (TPAMI 2026 · similarity 0.86)
- Adversarial Training on Purification (AToP): Advancing Both Robustness and Generalization (ICLR 2024 · similarity 0.86)
- Taming the Tail in Class-Conditional GANs: Knowledge Sharing via Unconditional Training at Lower Resolutions (CVPR 2024 · similarity 0.86)
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.
