Machine unlearning aims to make an AI model forget a previously learned concept. While generative unlearning has progressed rapidly, the unintended degradation of semantically related concepts that should remain—referred to as interference—remains poorly defined and inconsistently measured.
I-CARE treats interference as a first‑class research object by supplying formal task definitions, quantitative metrics, and a standardized reporting template. Rather than proposing a new benchmark or a specific unlearning algorithm, the framework enables systematic, reproducible measurement of interference across diverse unlearning scenarios.
The core components are:
- Task definition: precise specification of the target concept to erase, the concepts to retain, and the construction of evaluation data;
- Metrics: three‑dimensional assessment covering generation quality, concept retention rate, and interference magnitude;
- Reporting template: unified tables and visualizations to ensure results are comparable and reproducible.
To demonstrate feasibility, the authors applied state‑of‑the‑art unlearning methods (e.g., differential‑privacy fine‑tuning, inverse distillation) on several popular text‑to‑image models using common datasets such as COCO and LAION. The experiments show that I-CARE can reveal distinct interference patterns for different algorithm‑concept combinations, highlighting cases where unlearning unintentionally drifts semantic meaning.
An open‑source implementation and a web‑based graphical interface are provided, allowing users to explore outcomes without writing code or employing specialized analysis tools.
Review