I-CARE Framework Measures Interference in Text-to-Image Model Unlearning

Summary

Researchers introduce I-CARE, a methodology for studying interference in generative machine unlearning for text-to-image models. It defines tasks, metrics, and reporting templates for analyzing cases where removing one concept unintentionally harms related concepts that should be retained. A feasibility demonstration with current algorithms and common datasets found meaningful interference patterns. The project includes open-source software and a web interface for exploring results without coding or specialized analysis tools.