New Framework Scales AI Audit to Detect Model Shortcuts
ReMoDEx moves image classifier explainability from individual sample inspection to dataset-wide strategy clustering.
ReMoDEx provides a scalable pipeline to audit how image classifiers make decisions across large datasets. While deep learning models often show high predictive accuracy, they frequently rely on irrelevant cues or device artifacts rather than task-relevant data. This framework automates the detection of these shortcuts by grouping individual relevance maps into a few global decision strategy clusters.
https://arxiv.org/abs/2607.06889
The system integrates local explainability methods—including GradCAM++, Integrated Gradients, and Layerwise Relevance Propagation—with a global module that summarizes thousands of predictions. This replaces the manual, sample-by-sample inspection of heatmaps with a systematic assessment of spatial relevance.
Testing the framework on a VGG16 classifier for lung imaging revealed a critical gap in standard metrics. Despite a test accuracy of 86.27% and an AUC of 0.9624, ReMoDEx identified two recurring decision strategies: one focused on the central thoracic region and another sensitive to borders and corners. Masked image validation confirmed that occluding these peripheral regions altered the model's confidence and predicted class.
This shift from local to global explainability allows for the systematic identification of shortcut learning that accuracy-based evaluations miss. It transforms model auditing into a scalable process for verifying that AI decisions are grounded in the correct features.