Exposing Image Classifier Shortcuts with Counterfactual Frequency (CoF) Tables

Image classifiers often rely on irrelevant visual patterns rather than the intended features: The paper documents cases where AI models use shortcuts like copyright watermarks to identify horses, snowy backgrounds for huskies, and ink markings to detect skin cancer—none of which are reliable indicators in real-world scenarios.

Current methods for detecting these shortcuts are impractical at scale: Existing explainability approaches require researchers to manually examine numerous individual explanations to spot problematic patterns, making it difficult to systematically audit AI systems before deployment.

CoF (Counterfactual Frequency) tables offer a scalable solution for auditing image classifiers: By aggregating explanations across many images into summary tables, the method allows researchers to quickly identify which visual shortcuts a model has learned without needing to inspect thousands of individual cases.

Semantic labeling of image regions is necessary for meaningful audits: The approach requires humans to label what different parts of images represent, which adds a step but enables the system to distinguish between legitimate features and problematic shortcuts in a way that's interpretable to non-technical stakeholders.

This work addresses a critical gap in responsible AI deployment: By making it feasible to detect shortcuts before models are used in high-stakes applications (like medical imaging), the method helps prevent real-world harms from AI systems that appear accurate but rely on unreliable patterns.