Suppress and Diversify: Refining Robust Pathways for Corruption Robustness

Model robustness against natural image corruptions is essential for safety-critical applications. While existing methods primarily focus on implicit representation learning, we provide the first systematic exploration of computational pathways to explicitly characterize internal robustness. We identify a progressive decay of robust features across network layers and establish a functional dependency between the prevalence of these features and model performance. To exploit these insights, we propose Suppress and Diversify (S&D), a non-intrusive refinement approach that enhances robustness by dynamically selecting robust pathways and diversifying them through symmetry-preserving transformations. S&D is architecture-agnostic, parameter-free, and incurs zero test-time overhead. Extensive evaluations across eight benchmarks demonstrate that S&D consistently improves performance across multiple vision tasks, diverse backbones, and complex real-world scenarios, highlighting its broad efficacy and scalability.

We designed a method to make image recognition systems more reliable under real-world distortions, such as noise, blur, or weather effects, which can cause standard neural networks to misidentify objects. Instead of only training models to resist these corruptions indirectly, we analyzed how internal processing pathways—specific sequences of neuron activations—contribute to robustness. We found that robust pathways are concentrated in early layers and gradually diminish deeper in the network. Leveraging this, we propose a training refinement called Suppress and Diversify (S&D). S&D identifies the most robust pathways during training and applies transformations that maintain their structural properties while increasing diversity. This approach requires no extra parameters or runtime costs and can be applied to any existing network. Experiments across classification, detection, and segmentation tasks show that S&D consistently improves performance under various corruptions and out-of-distribution scenarios, demonstrating a practical and general strategy to strengthen neural networks against common real-world image degradations.