What Does It Mean to Forget a Person? Individual-Level Unlearning in Vision-Language Models
IDUnlearn-Bench finds vision-language models can drop a person's attributes yet still identify and reconstruct them.
Researchers introduce IDUnlearn-Bench, a benchmark for erasing a specific person from vision-language models rather than isolated attributes. Each individual is represented as linked multimodal evidence, scored on attribute access, identity access, identity binding, and identity reconstruction. Tests of representative VLMs and unlearning methods find that attribute-centric forgetting often leaves identity knowledge intact and can be non-monotonic. Models may suppress selected facts while still identifying, linking, or reconstructing the person.
- IDUnlearn-Bench tests attribute access, identity access, binding, and reconstruction.
- Attribute forgetting often leaves identity-level knowledge intact.
- Suppressing one risk can increase another, so unlearning is non-monotonic.
From the original article
1 min read at arxiv.orgErasing individual identities from Vision-Language Models (VLMs) is uniquely challenging because personal data is entangled across modalities rather than stored as isolated attributes.
By Xiongtao Sun · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.