SensX: Model-Agnostic Local Feature Attribution via Calibrated Global Sensitivity Analysis

Local feature attribution is a standard tool for auditing and debugging deep learning predictions, but existing attribution methods are not designed for systems that chain pretrained, frozen, or API-only modules. Gradient-based methods such as Integrated Gradients require an end-to-end computational graph that may be unavailable. Perturbation-based methods such as KernelSHAP require a reference input or background distribution whose choice can substantially alter attributions and may not be defensible for composite pipelines. We present SensX, a local attribution method that treats the model as a black box and replaces arbitrary design choices with interpretable, application-grounded parameters. SensX adapts Morris-style coordinate walks from global sensitivity analysis to local attribution. It requires no access to model internals, training data, or arbitrary reference inputs. We validate SensX across four case studies, each targeting a distinct limitation of existing methods. On a synthetic benchmark where ground-truth relevant features vary per input, SensX reaches 95% top-2 attribution accuracy versus 58% for the best KernelSHAP/Integrated Gradients variant. On a ViT with > 150,000 pixel-channel features, SensX produces spatially coherent maps and exposes systematic intra-patch bias where KernelSHAP is infeasible and Integrated Gradients yields task-irrelevant attributions. On single-cell classifiers with unstructured gene-expression features, SensX attains the lowest top- $k$ perturbation AUC. On a composite spatial transcriptomics system where neither method is applicable, SensX reveals reliance on preprocessing grid artifacts and a bias toward low-staining regions.