The Shape of Noise: Layer-Wise Perturbation Profiles for Diagnosing Vision Robustness
Abstract
As vision models are increasingly applied across diverse applications, the need for explainable and robust architectures continues to grow. Existing corruption robustness benchmarks, such as CIFAR-10-C, reduce model behavior to aggregate metrics like mean Corruption Error, obscuring how perturbations are amplified, suppressed, or transformed within a network. We introduce the Perturbation Evaluation Framework (PEF), a layer-wise diagnostic framework that decomposes a model’s response to input corruption and reveals architecture-dependent amplification and suppression signatures. Through experiments including intermediate-layer perturbation injection and profile-guided stabilization, we demonstrate that PEF complements aggregate benchmarks by providing interpretable layer-wise diagnostics for analyzing and targeting robustness behavior.