jiamingzz/geo_6k

---
license: mit
task_categories:
- image-classification
- visual-question-answering
language:
- en
tags:
- privacy
- adversarial-attack
- geographic-reasoning
- multimodal
size_categories:
- 1K<n<10K
---

# GeoPrivacy-6K

**[Project Page](https://jiamingzz94.github.io/reasonbreak/)** | **[Paper](https://arxiv.org/abs/2512.08503)** | **[Code](https://github.com/jiamingzhang94/ReasonBreak)**

## Introduction

**GeoPrivacy-6K** is a specialized dataset comprising **6,341 ultra-high-resolution images** ($\ge$ 2K resolution) designed to study and defend against reasoning-based privacy threats.

It was introduced in the paper **"Disrupting Hierarchical Reasoning: Adversarial Protection for Geographic Privacy in Multimodal Reasoning Models"** to train the **ReasonBreak** adversarial framework.

### Data Sources
The images are carefully curated and filtered from three established high-quality vision datasets to ensure diverse coverage of urban, natural, and aesthetic scenes:
* **HoliCity**: Urban environments with rich architectural details.
* **Aesthetic-4K**: Diverse high-quality scenes with strong composition.
* **LHQ (Landscapes HQ)**: Ultra-high-resolution natural landscapes.

### Privacy Note
This dataset focuses on *geographic* privacy. **It does not contain personally identifiable information (PII).** Sensitive elements such as human faces or license plates have been processed or anonymized in the original source datasets.

## Dataset Highlights

* **Hierarchical Conceptual Annotations**: Unlike traditional geolocation datasets that rely on GPS coordinates, GeoPrivacy-6K provides **multi-level visual concept annotations** (e.g., *"Gothic architecture"*, *"Deciduous forest"*) spanning Continental, National, City, and Local levels. This enables models to learn the *reasoning logic* behind location inference rather than just memorizing coordinates.
* **Fine-Grained Details**: All images maintain ultra-high resolution to preserve subtle cues (signage, vegetation patterns, architectural styles) that modern Multimodal Large Reasoning Models (MLRMs) exploit.

## Dataset Structure

* **Images**: Located in the root directory (zipped).
* **Annotations**: `location_analysis_fixed.jsonl` contains the reasoning chains, hierarchical concepts, and spatial bounding boxes for each image.

## Usage

This dataset is primarily designed for training the **ReasonBreak** generator.

Please refer to the [GitHub Repository](https://github.com/jiamingzhang94/ReasonBreak) for:
* Data loading scripts.
* Training instructions.

## Citation

If you use this dataset, please cite our paper:

```bibtex
@article{zhang2025reasonbreak,
title={Disrupting Hierarchical Reasoning: Adversarial Protection for Geographic Privacy in Multimodal Reasoning Models},
author={Zhang, Jiaming and Wang, Che and Cao, Yang and Huang, Longtao and Lim, Wei Yang Bryan},
journal={arXiv preprint arXiv:2512.08503},
year={2025}
}