MFPP: Morphological Fragmental Perturbation Pyramid for Black-Box Model Explanations

Black-box explanations don't require access to a model's internal workings: MFPP can identify which parts of an image a deep neural network uses to make decisions without needing to understand how the model is built internally, making it applicable to proprietary or commercial AI systems where internal details aren't available.

Visual explanations can be generated by strategically hiding image regions: The method works by masking different parts of images at multiple scales and observing how predictions change, creating "saliency maps" that show which pixels matter most—similar to how you might cover parts of a photo to see what's essential for recognizing an object.

Multi-scale analysis captures both fine and broad details: By examining image fragments at different sizes (a pyramid structure), MFPP better captures both small, detailed features and larger contextual patterns, outperforming simpler methods that only look at one level of detail.

This approach addresses a critical transparency gap in high-stakes applications: The paper emphasizes that DNNs used in medical diagnosis and autonomous driving need explainability; MFPP provides a practical tool to make these "black-box" decisions more interpretable to doctors, engineers, and regulators.

Performance improvements are demonstrated across multiple models and datasets: The authors show quantitatively that MFPP meets or exceeds existing explanation methods, suggesting it's a robust approach rather than a narrow solution for specific cases.