Research map / Image generation / GANs
ConfounderGAN: Protecting Image Data Privacy with Causal Confounder
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Use ConfounderGAN instead of EMN for in-distribution encryption | Image classification data · SVHN, CIFAR10, ImageNet25, BloodMNIST, Keratitis, ISIC | test accuracy (%): SVHN: 32.6 -> 25.4; CIFAR10: 19.1 -> 16.9; ImageNet25: 34.0 -> 26.5; BloodMNIST: 35.4 -> 29.7; Keratitis: 55.4 -> 35.6; | Helped |
| Use ConfounderGAN noise rather than EMN noise | Transferred encryption; evaluated on ResNet50 · SVHN, CIFAR10, ImageNet25, BloodMNIST, Keratitis, ISIC | test accuracy (%): SVHN: 10.6 -> 10.1; CIFAR10: 17.9 -> 11.3; ImageNet25: 12.3 -> 8.4; BloodMNIST: 41.7 -> 31.4; Keratitis: 34.4 -> 26.6; I | Helped |
| Train with the FA augmentation policy | CIFAR10 encrypted training data · CIFAR10; FA | test accuracy (%): 38.4 -> 39.3 | Hurt |
| Apply the other listed CIFAR10 augmentations rather than EMN | ConfounderGAN-encrypted data · CIFAR10; no augmentation, RHF, RC, RHF+RC, CutOut, MixUp, CutMix | test accuracy (%): No augmentation: 17.4 -> 11.2; RHF: 17.9 -> 11.9; RC: 18.0 -> 14.8; RHF+RC: 18.3 -> 14.9; CutOut: 18.2 -> 15.0; MixUp: 2 | Helped |
| Use ConfounderGAN for out-of-distribution encryption instead of DeepConfuse | CIFAR10 · CIFAR10; p=50% in-distribution plus out-of-distribution encrypted data | test accuracy (%): 71.3% -> 60.2% | Helped |
| Train a denoiser against ConfounderGAN encryption | ConfounderGAN-encrypted CIFAR10 · CIFAR10; adaptive denoiser | test accuracy (%): 11.9% -> 77.9% | Hurt |
Closest papers by meaning
- PATE-GAN: Generating Synthetic Data with Differential Privacy Guarantees (ICLR 2019 · similarity 0.87)
- RPGen: Robust and Differentially Private Synthetic Image Generation (AAAI 2026 · similarity 0.86)
- Reducing information dependency does not cause training data privacy. Adversarially non-robust features do. (ICLR 2026 · similarity 0.85)
- Protecting Facial Privacy: Generating Adversarial Identity Masks via Style-Robust Makeup Transfer (CVPR 2022 · similarity 0.85)
- Auditing Privacy Defenses in Federated Learning via Generative Gradient Leakage (CVPR 2022 · similarity 0.85)
- Balancing Privacy and Performance: A Many-in-One Approach for Image Anonymization (AAAI 2025 · similarity 0.85)
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.