Toward Faithful Segmentation Attribution via Benchmarking and Dual-Evidence Fusion

Post-hoc feature attribution methods are widely deployed in safety-critical vision systems, yet their stability under realistic input perturbations remains poorly characterized. Existing metrics evaluate explanations primarily under additive noise, collapse stability to a single scalar, and fail to condition on prediction preservation, conflating explanation fragility with model sensitivity. We introduce the Feature Attribution Stability Suite (FASS), a benchmark that enforces prediction-invariance filtering, decomposes stability into three complementary metrics: structural similarity, rank correlation, and top-k Jaccard overlap-and evaluates across geometric, photometric, and compression perturbations. Evaluating four attribution methods (Integrated Gradients, GradientSHAP, Grad-CAM, LIME) across four architectures and three datasets-ImageNet-1K, MS COCO, and CIFAR-10, FASS shows that stability estimates depend critically on perturbation family and prediction-invariance filtering. Geometric perturbations expose substantially greater attribution instability than photometric changes, and without conditioning on prediction preservation, up to 99% of evaluated pairs involve changed predictions. Under this controlled evaluation, we observe consistent method-level trends, with Grad-CAM achieving the highest stability across datasets.

Attribution methods in explainable AI rely on upsampling techniques that were designed for natural images, not saliency maps. Standard bilinear and bicubic interpolation systematically corrupts attribution signals through aliasing, ringing, and boundary bleeding, producing spurious high-importance regions that misrepresent model reasoning. We identify that the core issue is treating attribution upsampling as an interpolation problem that operates in isolation from the model's reasoning, rather than a mass redistribution problem where model-derived semantic boundaries must govern how importance flows. We present Universal Semantic-Aware Upsampling (USU), a principled method that reformulates upsampling through ratio-form mass redistribution operators, provably preserving attribution mass and relative importance ordering. Extending the axiomatic tradition of feature attribution to upsampling, we formalize four desiderata for faithful upsampling and prove that interpolation structurally violates three of them. These same three force any redistribution operator into a ratio form; the fourth selects the unique potential within this family, yielding USU. Controlled experiments on models with known attribution priors verify USU's formal guarantees; evaluation across ImageNet, CIFAR-10, and CUB-200 confirms consistent faithfulness improvements and qualitatively superior, semantically coherent explanations.

Attribution methods are widely used to characterize the evidence underlying model predictions, yet their potential to improve model behavior remains underexplored. Attribution inconsistency under label-preserving geometric transformations may indicate transformation-sensitive evidence reliance, motivating attribution regularization. However, such supervision is valid only when attribution faithfully reflects the evidence driving predictions. Existing self-supervised methods typically align gradient-based maps such as Grad-CAM, whose limited faithfulness means that attribution consistency need not imply consistency of the underlying decision process, leaving transformation robustness unresolved. We propose an annotation-free attribution regularization framework based on submodular search over image regions. By measuring how candidate subsets affect model outputs, the search extracts compact, class-discriminative evidence as search-derived supervision. We further introduce a submodular ranking loss with path-consistency and termination-alignment terms that respectively align spatially corresponding candidate rankings along paired search trajectories and encourage the transformed trajectory to satisfy the stopping criterion at the target terminal step. The loss provides a differentiable surrogate for regularizing both final attributions and the otherwise discrete evidence-selection process. Experiments on ImageNet-100 show that our method substantially improves attribution stability, Insertion, and Deletion on ViT-B/16 with only a 0.28-point accuracy drop, with similar gains on ViT-L/16. On ImageNet-1K, it improves transformed-input accuracy on ResNet-50 and ConvNeXt-B while limiting the clean-accuracy drop to 0.30 points, demonstrating more consistent evidence reliance with minimal performance loss. Code will be released soon.