corruption simulation

Real-world distribution shifts—such as weather and illumination variations—severely degrade the robustness of deep learning models. However, collecting diverse, real-world degraded data is prohibitively expensive, prompting widespread reliance on synthetic degradation; yet its fidelity in reflecting real-world degradation effects remains unclear. Method: We construct the largest cross-domain (real vs. synthetic) semantic segmentation corruption benchmark to date, built upon Cityscapes and other datasets using the CorruptIO toolkit. We systematically evaluate 12 corruption types across multiple models and metrics (mIoU, RankCorr). Contribution/Results: We discover, for the first time, a strong correlation (ρ = 0.89) between model performance under real and synthetic corruptions. We further propose a corruption-type-level correlation analysis framework to characterize the applicability boundaries of synthetic degradation. All evaluation code, protocols, and benchmarks are publicly released to advance standardized robustness assessment.

Deep learning models exhibit insufficient robustness against adversarial perturbations and common image corruptions, undermining their reliability in real-world deployment. To address this, we propose an active robustness verification strategy that leverages the training set itself: by performing local robustness analysis, our method automatically identifies “weakly robust” samples—serving as early, interpretable indicators of model vulnerability—and enables targeted robustness enhancement. Unlike conventional passive paradigms that rely solely on perturbed test sets for robustness evaluation, ours is the first to repurpose training data for robustness diagnostics. We integrate adversarial perturbation injection with diverse natural corruption tests. Extensive experiments on CIFAR-10, CIFAR-100, and ImageNet demonstrate that our strategy significantly improves model robustness against both attacks and corruptions (average gain of +8.2%) while enhancing the sensitivity and interpretability of reliability assessment.

This work addresses the challenge of evaluating spatial robustness of dense vision models (e.g., semantic segmentation) under localized natural and adversarial corruptions in safety-critical applications such as autonomous driving. We formally define and quantify “spatial robustness” for the first time, and propose a region-aware multi-attack adversarial analysis framework that transcends conventional single-point perturbation paradigms. Our analysis reveals complementary vulnerabilities between CNNs and Transformers: Transformers exhibit higher resilience to natural corruptions but greater susceptibility to adversarial attacks, whereas CNNs show the opposite pattern. Leveraging localized corruption modeling, multi-objective region-based attacks, and a novel spatial robustness metric, we establish a new benchmark. Extensive evaluation across 15 models demonstrates that our integrated defense strategy significantly enhances comprehensive robustness and reliability against both natural and adversarial spatial threats.

Existing optical flow, scene flow, and stereo vision benchmarks evaluate only accuracy, lacking systematic assessment of robustness to realistic image degradations (e.g., noise, rain, fog, blur). Method: We introduce the first unified robustness benchmark for these three tasks, covering 20 real-world degradation types. We propose a novel time-disparity-depth-consistent degradation generation paradigm and a new robustness metric, elevating robustness to a core evaluation dimension on par with accuracy. Leveraging the high-resolution Spring dataset, we release 20,000 high-fidelity degraded images. Contribution/Results: We establish a dual-axis evaluation framework jointly measuring accuracy and robustness. Extensive experiments reveal weak correlation between accuracy and robustness, and substantial performance variation across degradation types—highlighting critical limitations of current models. This benchmark provides foundational infrastructure and actionable insights for developing robust vision models.

Existing DNN robustness auditing methods are constrained to predefined, isolated image distortions and thus fail to characterize complex, coupled degradations arising in real-world imaging. This work breaks from conventional correlation-based testing paradigms by introducing causal inference into model robustness auditing for the first time. We formalize the imaging process as a causal graph and quantify the causal effects of low-level factors—such as illumination variations, sensor noise, and motion blur—on model performance via counterfactual reasoning and observational-domain-driven causal effect estimation. Evaluated across diverse visual tasks and domains (natural vs. rendered images), our approach is both interpretable and transferable. It significantly improves the precision of robustness failure localization and enhances predictive capability for deployment risks under realistic degradation conditions.

Clinical CT images often suffer from degradations such as noise, low resolution, and contrast variations, which significantly impair segmentation performance. To address this challenge, this work proposes RAMP, a novel framework that systematically integrates clinically motivated multi-degradation augmentation strategies. Specifically, RAMP enhances model robustness by combining anatomically constrained spatial perturbations, CT-specific intensity transformations, and stochastic combinations of multiple degradations, building upon the nnU-Net architecture. Evaluated on both a five-organ dataset and Abdomen1K, RAMP achieves average Dice scores of 0.753 and 0.789, respectively, on degraded images, reducing the robustness gap to only 0.064 and 0.070. These results demonstrate that RAMP substantially mitigates segmentation collapse under severe image degradation.

This work proposes a novel framework based on adaptive feature fusion and contrastive learning to address the limited generalization of existing methods in complex scenarios. By dynamically integrating multi-scale semantic information and incorporating cross-sample consistency constraints, the approach significantly enhances model robustness under distribution shifts. Extensive experiments demonstrate that the proposed method consistently outperforms state-of-the-art models across multiple benchmark datasets, with particularly notable gains in low-resource and long-tailed settings. Beyond offering a new perspective for improving model generalization, this study also releases the associated code and pre-trained models to facilitate future research.

Deep learning models for medical imaging often suffer performance degradation in new clinical settings due to distribution shifts—such as changes in imaging devices, patient populations, or acquisition protocols—yet existing stress tests rely on unrealistic perturbations that poorly reflect real-world robustness. This work proposes the first counterfactual stress-testing framework based on causal generative models, which intervenes on variables like scanner type or patient sex to synthesize clinically plausible “what-if” images that preserve anatomical structure while realistically simulating target distribution shifts. Experiments on chest X-ray and mammography datasets demonstrate that this approach more accurately predicts the direction, magnitude, and relative ranking of model performance changes under real out-of-domain conditions compared to conventional perturbation methods, substantially improving the correlation between stress-test outcomes and actual model robustness.

This study investigates the relationship between the robustness of neural networks under random input perturbations and their prediction accuracy, measured by mean squared error (MSE). To address this, the work proposes an efficient, computable black-box robustness metric that, without requiring access to internal model architecture, provides a high-probability upper bound on the network’s MSE over an entire dataset under a given perturbation. The method innovatively introduces robustness curves, enabling systematic comparison and analysis of robustness across different datasets. Experimental evaluations on multiple real-world datasets demonstrate that the proposed approach accurately quantifies and effectively captures a model’s sensitivity to input noise, offering a practical tool for assessing robustness in diverse settings.

This work addresses the critical gap in understanding whether synthetic images are truly interchangeable with real ones in model training and the absence of systematic evaluation frameworks to ensure their safe and effective use. The study systematically quantifies discrepancies between synthetic and real images across three dimensions: high-dimensional feature distributions, low-level statistical properties in color space, and model training dynamics. Building on these insights, the authors propose a pre-evaluation metric for synthetic data of unknown quality and a safety-aware data fusion strategy for training. Experiments demonstrate that carefully calibrated mixing ratios and integration methods of synthetic and real data can substantially enhance model performance and robustness, thereby offering both theoretical grounding and practical guidance for the reliable deployment of synthetic data in machine learning pipelines.