DEXTER: Diffusion-Guided EXplanations with TExtual Reasoning for Vision Models

Data-free explanations improve privacy and accessibility: DEXTER can explain how image classifiers make decisions without needing access to the original training data or labeled examples, making it practical for auditing AI systems where data sharing isn't possible due to privacy concerns.

The framework identifies hidden biases in visual AI: By generating synthetic images that activate classifiers and analyzing them with language models, DEXTER can uncover class-specific decision patterns and biases—such as spurious correlations—that might otherwise remain hidden from developers and users.

Natural language explanations make AI more understandable: Rather than producing technical visualizations, DEXTER generates human-readable text reports describing why a classifier makes decisions, making it easier for non-technical stakeholders to understand and evaluate whether a model's reasoning is trustworthy.

Performance validated across multiple real-world scenarios: Testing on standard datasets (ImageNet, Waterbirds, CelebA, FairFaces) and including a user study demonstrates that DEXTER's explanations are both accurate and interpretable in practical applications, not just theoretical exercises.

Combines existing AI tools for transparency: By leveraging diffusion models (for image generation) and large language models (for reasoning), DEXTER shows how combining different AI capabilities can create new tools for responsible AI—potentially making model auditing more accessible to organizations without specialized expertise.