Research map / Adversarial robustness and security / Certified and empirical robustness
Towards Robustness to Natural Variations and Distribution Shift (Student Abstract)
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Compares standard, adversarial, augmentation, and model-based/hybrid robust training methods on SVHN under brightness distribution shift, using a small CNN and a trade-space framing.
Problem
Robustness of ML models to natural variations and distribution shifts (e.g., brightness changes), where it is unclear which data-variation training approach works best.
Approach
Presents a trade space that places robust training methods by the type of data variation (additive vs. natural) and whether variations are task-independent or optimized for the task, and introduces three hybrid model-based methods (MDAT, MRAT, AugMix-RT) combining learned variation models with PGD or AugMix.
Key results
- Standard and AT models drop to 44.3 and 50.7 accuracy under high brightness shift, versus about 77-83 for augmentation and model-based methods.
- AugMix achieves 82.6 accuracy under high brightness with calibration error 7.8, while Standard has calibration error 47.8 on the same shift.
- MAT is reported as more stable (lowest calibration error 5.5 and 80.3 high-brightness accuracy), though the authors state no clear winner overall.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DataAug instead of no data variation | Standard · SVHN medium brightness (MB) | accuracy: 67.4 +/- 4.2 -> 86.6 +/- 1.0 | Helped |
| Use AugMix instead of no data variation | Standard · SVHN high brightness (HB) | accuracy: 44.3 +/- 4.3 -> 82.6 +/- 1.8 | Helped |
| Use adversarial training instead of standard training | Standard · SVHN high brightness (HB) | accuracy: 44.3 +/- 4.3 -> 50.7 +/- 7.5 | Helped |
| Train with samples from a learned variation model (MDA) rather than standard training | Standard · SVHN high brightness (HB) | accuracy: 44.3 +/- 4.3 -> 78.4 +/- 2.1 | Helped |
| Use PGD to optimize learned-model variations (MAT) rather than random learned-model samples (MDA) | MDA · SVHN high brightness (HB) | accuracy: 78.4 +/- 2.1 -> 80.3 +/- 0.2 | Helped |
| Use PGD to optimize learned-model variations (MAT) rather than random learned-model samples (MDA) | MDA · SVHN high brightness (HB) | calibration error: 7.7 +/- 1.3 -> 5.5 +/- 0.7 | Helped |
| Use AugMix augmentation instead of learned variation-model sampling | MDA · SVHN high brightness (HB) | accuracy: 78.4 +/- 2.1 -> 82.6 +/- 1.8 | No clear effect |
| Replace MRT's learned variation model with AugMix | MRT · SVHN high brightness (HB) | accuracy: 79.8 +/- 0.4 -> 82.7 +/- 0.7 | No clear effect |
| Use the AugMix-RT hybrid instead of MRT | MRT · SVHN high brightness (HB) | calibration error: 7.0 +/- 0.9 -> 5.8 +/- 1.4 | Helped |
| Use the MDAT hybrid instead of MAT | MAT · SVHN high brightness (HB) | accuracy: 80.3 +/- 0.2 -> 78.2 +/- 1.2 | Hurt |
| Use the MRAT hybrid instead of MAT | MAT · SVHN high brightness (HB) | accuracy: 80.3 +/- 0.2 -> 76.9 +/- 0.8 | Hurt |
Surprising findings
- Learned variation models (MDA) showed no noticeable advantage over simple analytical transforms such as AugMix.
- Model-based and hybrid methods had lower variance and lower calibration error than DataAug and AugMix in some cases, but the paper reports no clear winner.
Limitations
- Experiments limited to SVHN and a small CNN with a single natural variation (brightness)
- No clear winner among methods; differences often within error bars
- Only one real-world-like application planned; medical imaging left to future work
Experimental setup
Datasets: SVHN MNIST
Benchmarks: SVHN brightness shift test sets (low, medium, high brightness)
Metrics: accuracy calibration error
Baselines: Standard AT DataAug AugMix MDA MRT MAT
Builds on: Madry adversarial training (PGD) AugMix Model-based robust training (Robey et al. 2020)
Closest papers by meaning
- Measuring Robustness to Natural Distribution Shifts in Image Classification (NeurIPS 2020 · similarity 0.89)
- Outlier-Robust Distributionally Robust Optimization via Unbalanced Optimal Transport (NeurIPS 2024 · similarity 0.87)
- Certifying Some Distributional Robustness with Principled Adversarial Training (ICLR 2018 · similarity 0.86)
- Stable Adversarial Learning under Distributional Shifts (AAAI 2021 · similarity 0.86)
- Distributionally Robust Optimization with Bias and Variance Reduction (ICLR 2024 · similarity 0.86)
- Tolerant Algorithms for Learning with Arbitrary Covariate Shift (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.