Hide and Seek: Black-Box Concept Attribution Without Model Weights or Labeled Data

Forum topic · 小凯 · 2026-10-13

Summary

Hide and Seek (HnS) is a new concept-level attribution method for black-box vision models introduced in arXiv paper 2610.13935 by Tokas, Nair, Jackson, and Kerner. Unlike existing concept-based explanation methods, HnS requires neither access to model weights nor large concept-annotated datasets, and it works for tasks beyond classification and regression. Given a concept and an image set, the method uses a vision-language model to retrieve concept-aligned images, localizes the concept within each image, masks the localized region, and measures attribution through the change in the black-box model's output between the original and masked images—directly probing model behavior and enabling local, image-level attribution. Experiments on CelebA and COCO-Stuff show better alignment with ground-truth attribution than prior methods. The authors also demonstrate applicability to image captioning and satellite image segmentation, tasks no existing concept attribution method could previously handle. This makes HnS a practical, model-agnostic tool for interpreting diverse computer vision systems.

Hide and Seek: HIDE Concepts to SEEK Explanations for Black Box Models

Field: Computer Vision Authors: Bhanu Tokas, Rahul Nair, Rockwell Jackson, Hannah Kerner Published: 2026-10-09 arXiv: 2610.13935

Key Points

Interpreting predictions of black-box vision models is a central challenge in machine learning. Concept-based attribution methods aim to explain models using human-understandable concepts.
Limitations of existing methods:
Only applicable to classification and regression models
Require access to model weights or large concept-annotated datasets
Most cannot provide image-level attribution
Proposed method — Hide and Seek (HnS):
Requires no access to model weights and no labeled concept datasets
Can interpret a wide range of vision models
Pipeline: given a concept and an image set, HnS uses a vision-language model to retrieve concept-aligned images, localizes the concept within each image, masks the localized region, and measures attribution via the change in the black-box model's output between the original and masked images
Directly measures model behavior, supporting local attribution
Evaluation:
Validated on CelebA and COCO-Stuff, showing better agreement with ground-truth attribution than existing methods
Demonstrated on image captioning and satellite image segmentation — tasks for which no existing concept attribution method is applicable

Original Abstract (Excerpt)

> Interpreting machine learning model predictions is a central challenge, particularly for black-box vision models. Concept-based attribution methods interpret models using human-understandable concepts. Existing concept-based methods can only interpret classification and regression models. They require access to model weights or large concept-annotated datasets, and most methods cannot provide image-level attribution. We propose Hide and Seek (HnS), a new method for concept-level attribution that does not require access to model weights or labeled concept datasets. HnS can interpret a wide range of vision models. Given a concept and image set, HnS retrieves concept-aligned images using a vision-language model, localizes the concept within each image, masks the localized region, and measures...

--- *Auto-collected on 2026-10-13*