Robust Evaluation of Attribution Maps with Morphology-Based Framework for Reliable Interpretability in Vision Models

• Autores: Stanchi, Oscar Agustín; Ronchetti, Franco; Quiroga, Facundo Manuel; Dal Bianco, Pedro Alejandro; Hasperué, Waldo; Rashwan, Hatem; Puig, Domènec
• Año de publicación: 2025
• Idioma: inglés
• Tipo de recurso: documento de conferencia
• Materia: Ciencias Informáticas; Attribution; Computer Vision; Deep Learning; Feature Importance; Morphological Operations; Interpretability; Quantitative Metrics
• Institución: Universidad Nacional de La Plata

Deep neural networks have achieved remarkable performance across computer vision tasks, but their opaque decision-making processes remain a significant barrier to trustworthy deployment in high-stakes domains. Attribution methods are a widely used approach in computer vision to generate importance maps to explain model behavior. However, to choose among the many attribution methods, researchers and practitioners typically choose methods in a subjective fashion, cherry picking results and lacking a robust evaluatuion of their performance. To address this critical gap, we introduce a morphology-based framework for rigorous assessment of attribution map robustness. Our approach establishes axiomatic safeguards against trivial explanation strategies and implements two complementary metrics: 1) Morph Erosion Score (MES), measuring preservation of critical features under morphological erosion; and 2) Least Important First (LIF) Score, evaluating tolerance to removal of unimportant pixels. Through comprehensive experiments on 1,143 ImageNet-1K images using ResNet-101, we demonstrate that our approach reveals previously obscured method characteristics: CB-RISE dominates spatial coherence (MES: 0.641), while Occlusion excels in importance ranking (LIF: 0.681). Crucially, trivial strategies fail catastrophically (LIF AUC ≤ 0.001), validating our axioms. By exposing inherent tradeoffs between attribution properties, our methodology provides actionable insights for selecting context-appropriate interpretability methods and advances progress toward certified-explainable vision systems.
Red de Universidades con Carreras en Informática
Instituto de Investigación en Informática