Diffusion Integrated Gradients: Controllable Path Generation for Flexible Feature Attribution

Attribution paths significantly impact explanation quality: The paper demonstrates that how we trace explanations from a baseline to an input substantially affects the quality of AI model explanations, challenging the common assumption that existing fixed-path methods are sufficient for understanding model decisions.

Traditional attribution methods produce noisy explanations: Existing path-based attribution approaches like Integrated Gradients rely on hand-crafted paths that often generate unclear or distorted explanations, limiting their reliability for real-world applications where stakeholders need trustworthy model interpretability.

Generative AI can improve explainability: By using diffusion models (a type of generative AI) to learn better paths for attribution, DiffIG achieves more perceptually aligned explanations—meaning the explanations better match human intuition about what features actually matter in model decisions.

Users can guide explanations at inference time: Unlike previous methods requiring fixed design choices upfront, DiffIG allows users to embed their own guidance during the explanation process, making XAI methods more flexible and adaptable to different stakeholder needs and use cases.

Better explanations support responsible AI deployment: By producing clearer, more controllable explanations of model predictions, this work contributes to building more trustworthy AI systems where organizations can better understand, audit, and justify their models' decisions to users and regulators.