Research map / Adversarial robustness and security / Certified and empirical robustness
Solving Inverse Problems With Deep Neural Networks – Robustness Included?
Martin Genzel, Jan Macdonald, Maximilian März
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 |
|---|---|---|---|
| train with noise jittering instead of noiseless data (ablation) | ItNet · MNIST compressed sensing (Scenario A2), adversarial noise | relative reconstruction error: error blows up at ~15% adversarial noise without jittering | Improved |
| train with noise jittering instead of noiseless data (ablation) | UNet · Radon ellipses (Scenario B2), 1% adversarial noise | reconstruction robustness | Improved |
| use original fastMRI vertical-line mask instead of radial mask | Tira and TV · fastMRI single-coil, 4-fold acceleration | adversarial robustness: undersampling ~23% vs ~17% | Hurt |
| use ItNet data-consistent iterative architecture instead of plain post-processing | ItNet · CS Gaussian measurements | robustness and accuracy | Improved |
| use Tiramisu instead of U-Net | Tiramisu · CS Gaussian measurements | reconstruction accuracy | Improved |
| make inversion layer learnable (fully-learned) (ablation) | UNetFL/ItNetFL · CS Gaussian measurements | reconstruction accuracy | Improved |
| fully-learned inversion layer instead of Fourier adjoint | fully learned schemes · Fourier ellipses (Scenario B1) | reconstruction accuracy | No clear effect |
| replace TV minimization with learned networks under statistical noise | NN-based methods · all case studies, mid to high noise | relative error | Improved |
| replace TV minimization with learned networks under adversarial noise | NN-based methods · all case studies | relative error | No clear effect |
| add a ConvNet classifier after reconstruction (compressed classification) | Rec + ConvNet · MNIST from Gaussian measurements | accuracy: drops to zero at some noise level | Hurt |
Closest papers by meaning
- Measuring Robustness in Deep Learning Based Compressive Sensing (ICML 2021 · similarity 0.95)
- Improving Robustness of Deep-Learning-Based Image Reconstruction (ICML 2020 · similarity 0.93)
- Robustness of Deep Learning for Accelerated MRI: Benefits of Diverse Training Data (ICML 2024 · similarity 0.87)
- The Perception-Robustness Tradeoff in Deterministic Image Restoration (ICML 2024 · similarity 0.87)
- SGD Jittering: A Training Strategy for Robust and Accurate Model-Based Architectures (ICML 2025 · similarity 0.87)
- A Restoration Network as an Implicit Prior (ICLR 2024 · similarity 0.87)
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.
