Decoupling Pixel Flipping and Occlusion Strategy for Consistent XAI Benchmarks

Occlusion strategies in AI explanations are inconsistent: The paper identifies that different methods for removing features from images (ranging from simple averaging to advanced AI-based inpainting) produce contradictory results when evaluating which features are most important to a model's decisions, undermining trust in these explanations.

Current benchmarking methods give conflicting rankings: Existing pixel-flipping evaluation methods rank explanation quality differently depending on whether they test by removing the most important features first or least important features first, making it impossible to reliably compare different explanation techniques.

A new reliability metric helps resolve disagreements: The researchers propose the "Reference-Out-of-Model-Scope" (R-OMS) score to measure whether occlusion strategies create unrealistic images that fool models, allowing systematic comparison of different approaches and grouping them by consistency.

Combining two opposing approaches yields stable results: By merging the two conflicting evaluation methods into a single "symmetric relevance gain" (SRG) measure, the authors achieve consistent rankings across 40 different occlusion strategies, providing a more reliable foundation for XAI benchmarking.

This work addresses a fundamental problem in responsible AI: By standardizing how we evaluate explanations, this research helps ensure that AI systems can be reliably audited and understood, which is essential for deploying trustworthy AI in high-stakes applications.