Research map / Adversarial robustness and security / Attacks on text and graphs
Deep Perturbation Learning: Enhancing the Network Performance via Image Perturbations
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 |
|---|---|---|---|
| Add DPL (augmentation, iterative) to standard training | VGG-16, ResNet-18, ResNet-50, PreActResNet-18, DenseNet-121, MobileNetV2, ViT-B · CIFAR-10 / CIFAR-100 / ImageNet classification | Top-1 Acc.: up to +0.98 (CIFAR-10), +2.01 (CIFAR-100), +0.71 (ImageNet) | Helped |
| Add DPL to standard training | ResNet-50 · ImageNet | Top-1 Acc.: +0.36 | Helped |
| Add DPL to standard training | Faster R-CNN, SSD300, EfficientDet-D0, MobileNetV2, RetinaNet · PASCAL VOC and COCO object detection | mAP: +0.3 to +0.5 (VOC); +0.3 and +0.2 (COCO) | Helped |
| Add DPL to EMANet and CFNet | EMANet (R-50, R-151), CFNet (R-101) · PASCAL VOC + SBD segmentation | mIoU: +0.29, +0.05, +0.24 | Helped |
| Adversarial perturbation methods (I-FGSM, PGD, IAT, AdvProp) used as training augmentation instead of DPL | ResNet-50 · CIFAR-100 and ImageNet classification | Top-1 Acc.: I-FGSM -7.41 (CIFAR-100), -4.26 (ImageNet); AdvProp (35 epochs) -1.97 / -2.19 | Hurt |
| Add DPL on top of MixUp and CutMix (ablation) | ResNet-50 · ImageNet classification | Top-1 Acc.: MixUp 77.33->77.51; CutMix 78.46->78.57 | Helped |
| Increase perturbation strength alpha from 0.0001 to 0.005 (single-iter DPL) (ablation) | PreActResNet-18 · CIFAR-10 | Top-1 Err.: error 4.92 baseline; 4.66 at alpha=0.005 (best); 4.84 at 0.01; 4.74 at 0.0001 | Mixed |
| Increase number of amended samples (epsilon) beyond about 4% of training set (ablation) | PreActResNet-18 with single-iteration DPL · CIFAR-10 | Top-1 Err.: error rate rebounds beyond 4% (best 4.63%) | Hurt |
| Iterate DPL more than about 3 rounds (ablation) | Single-model CIFAR-10 training · CIFAR-10 | test loss / expected risk | Hurt |
| Use replacement instead of augmentation for amended set (ablation) | DPL on CIFAR-10 · CIFAR-10 | test error | Mixed |
Closest papers by meaning
- Defense Against Universal Adversarial Perturbations (CVPR 2018 · similarity 0.91)
- Wide Two-Layer Networks can Learn from Adversarial Perturbations (NeurIPS 2024 · similarity 0.88)
- Contrasting Adversarial Perturbations: The Space of Harmless Perturbations (AAAI 2025 · similarity 0.87)
- Towards Transferable Adversarial Attacks with Centralized Perturbation (AAAI 2024 · similarity 0.86)
- Adversarial Perturbations Are Formed by Iteratively Learning Linear Combinations of the Right Singular Vectors of the Adversarial Jacobian (ICML 2025 · similarity 0.86)
- Improving robustness to corruptions with multiplicative weight perturbations (NeurIPS 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.