Research map / Adversarial robustness and security / Attacks on text and graphs
Improving the Robustness of Deep Neural Networks via Stability Training
Check 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 stability training to deep ranking features | Inception-based deep ranking · Similar-image ranking, Original | ranking score @top-30: 7,312 -> 7,368 | Helped |
| Add stability training to deep ranking features | Inception-based deep ranking · Similar-image ranking, JPEG-50, thumb-30k, and crop-10 | ranking score @top-30: 7,286 -> 7,360; 7,160 -> 7,172; 7,298 -> 7,322 | Helped |
| Add stability training to ranking features | Inception-based deep ranking · Near-duplicate detection, thumbnail, JPEG, and crop distortions | recall at fixed precision: +1.0% at 99.5% precision; +3.0% at 98% precision | Helped |
| Add stability training to ranking features | Inception-based deep ranking · JPEG-50 near-duplicate pairs | feature-distance distribution below 0.1: 76% -> 86% | Helped |
| Add stability training to Inception | ImageNet classification · ImageNet validation, Original | precision @top-5: 93.9% -> 93.6% | No clear effect |
| Add stability training to Inception | ImageNet classification · ImageNet validation, JPEG-50 and JPEG-10 | precision @top-5: 92.4% -> 92.7%; 83.0% -> 88.3% | Helped |
| Add stability training to Inception | ImageNet classification · ImageNet validation, Original | precision @top-1: 77.8% -> 77.9% | No clear effect |
| Add stability training to Inception | ImageNet classification · ImageNet validation, JPEG-50 and JPEG-10 | precision @top-1: 75.1% -> 75.7%; 61.1% -> 67.9% | Helped |
| Train on Gaussian-perturbed copies using the original task loss (ablation) | Deep ranking network · Image ranking | Triplet ranking score @top-30: 7,312 -> 6,300 -> 5,065 | Hurt |
| Fine-tune only the final fully-connected layers rather than stability-training the whole network from the beginning (ablation) | Inception-based networks · Classification and ranking tasks | model performance | No clear effect |
Closest papers by meaning
- Instant Video Models: Universal Adapters for Stabilizing Image-Based Networks (NeurIPS 2025 · similarity 0.89)
- Zero Stability Well Predicts Performance of Convolutional Neural Networks (AAAI 2022 · similarity 0.88)
- On the Stability of Graph Convolutional Neural Networks: A Probabilistic Perspective (NeurIPS 2025 · similarity 0.87)
- Stability and Generalization of Adversarial Training for Shallow Neural Networks with Smooth Activation (NeurIPS 2024 · similarity 0.86)
- Verification of Neural Networks Against Convolutional Perturbations via Parameterised Kernels (AAAI 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.