Research map / Probabilistic and scientific modeling / Variational and generative sampling
Semantic Perturbations With Normalizing Flows for Improved Generalization
Oguz Kaan Yüksel, Sebastian U. Stich, Martin Jaggi, Tatjana Chavdarova
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 adversarial latent attacks via normalizing flow | ResNet-18 standard training with common DA · CIFAR-10 (5% train) | test accuracy: 64.1 -> 80.4 | Improved |
| add randomized latent attacks via normalizing flow | ResNet-18 standard training with common DA · CIFAR-10 (5% train) | test accuracy: 64.1 -> 70.1 | Improved |
| add adversarial latent attacks | ResNet-18 · CIFAR-10 (full train set) | test accuracy: 95.2 -> 96.6 | Improved |
| use VAE-GAN on-manifold perturbation instead of none | ResNet-18 · CIFAR-10 (5% train) | test accuracy: 64.1 -> 58.9 | Hurt |
| use Adversarial-LA instead of Mixup | ResNet-18 · CIFAR-10 (5% train) | test accuracy: 73.4 -> 80.4 | Improved |
| add Adversarial-LA with flow trained on CIFAR-10 (transfer) | ResNet-18 · CIFAR-100 (10% train) | test accuracy: 36.4 -> 45.0 | Improved |
| train flow on CIFAR-10 instead of SVHN itself (ablation) | ResNet-18 with Adversarial-LA · SVHN (5% train) | test accuracy: 86.9 -> 90.5 | Improved |
| use latent-space attack instead of image-space PGD for training | ResNet-18 · CIFAR-10 (full train set) | test accuracy: 86.3 -> 96.6 | Improved |
| add Adversarial-LA training | ResNet-18 vs Adversarial-LA attack · CIFAR-10 | accuracy under attack: 58.8 -> 76.4 | Improved |
| add image-space PGD training | ResNet-18 vs Adversarial-LA attack · CIFAR-10 | accuracy under attack: 58.8 -> 36.2 | Hurt |
| add Adversarial-LA without any data augmentation | ResNet-18 · CIFAR-10 (5% train, no DA) | test accuracy: 49.8 -> 58.2 | Improved |
| use VAE-GAN without data augmentation | ResNet-18 · CIFAR-10 (5% train, no DA) | test accuracy: 49.8 -> 49.4 | No clear effect |
Closest papers by meaning
- Perturbation-Induced Linearization: Constructing Unlearnable Data with Solely Linear Classifiers (ICLR 2026 · similarity 0.88)
- Learning perturbation sets for robust machine learning (ICLR 2021 · similarity 0.87)
- Guided Perturbations: Self-Corrective Behavior in Convolutional Neural Networks (ICCV 2017 · similarity 0.87)
- Data-Efficient Augmentation for Training Neural Networks (NeurIPS 2022 · similarity 0.87)
- API-Net: Robust Generative Classifier via a Single Discriminator (ECCV 2020 · similarity 0.86)
- Understanding Deep Networks via Extremal Perturbations and Smooth Masks (ICCV 2019 · 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.
