Derivative-Free Diffusion Manifold-Constrained Gradient for Unified XAI

Black-box explainability is now possible: Unlike traditional gradient-based explanation methods that require full access to a model's internal workings, FreeMCG only needs access to the model's outputs, making it applicable to real-world scenarios where companies keep their AI systems proprietary.

Explanations are now more realistic and trustworthy: The method addresses a fundamental problem where previous explanation techniques produced unrealistic images that don't actually exist in the real world. By constraining explanations to stay on the "data manifold" (the space of realistic images), FreeMCG generates explanations that better align with how humans perceive images and are more faithful to what the model actually learned.

Explanations are more resistant to manipulation: Traditional gradient-based methods can be fooled by adversarial attacks—small, intentional perturbations designed to trick the explanation system. FreeMCG's derivative-free approach makes it harder to manipulate in this way, providing more reliable explanations that stakeholders can trust.

One method can do multiple explanation tasks: FreeMCG unifies two previously separate explanation approaches—counterfactual explanations (showing what would need to change for a different prediction) and feature attribution (showing which parts of an image matter most). This unified approach could simplify how organizations implement explainability tools.