ICLR 2026: ReasonBreak

Disrupting Hierarchical Reasoning: Adversarial Protection for Geographic Privacy in Multimodal Reasoning Models

Project Page | Paper | Dataset | Model

Official implementation of ReasonBreak, an adversarial framework designed to protect geographic privacy against Multimodal Large Reasoning Models (MLRMs).

🛠️ Installation

conda create -n reasonbreak python=3.10
conda activate reasonbreak
pip install -r requirements.txt

📂 Data Preparation

Please organize the directory as follows:

ReasonBreak/
├── data/
│   ├── embedding_bank.pth        # Download from Model repo
│   ├── json/
│   │   ├── location_analysis_fixed.jsonl  # For training
│   │   └── cot_full.json                  # For inference
│   └── images/
│       ├── geo_6k/               # [Optional] Training Set
│       └── dox/                  # [Required] Evaluation Set
├── checkpoints/
│   └── weight.pth                # ReasonBreak Weights
└── ...

Download Links

• Datasets:
  • GeoPrivacy-6K (Training): Optional. Download only if you intend to train the model from scratch.
  • DoxBench (Evaluation): Required. Necessary for running the evaluation pipeline.
• Model Weights:
  • ReasonBreak: Download weight.pth and place it in checkpoints/.

API Configuration

Create a .env file in code/experiment/ for evaluation:

OPENAI_API_KEY=sk-...
GOOGLE_API_KEY=AIza...

🚀 Quick Start

Run the unified script run.sh. Choose ONE of the following options:

Option 1: Full Pipeline (Recommended)

Generate adversarial samples and immediately evaluate performance.

bash run.sh --mode adv --step all --dox_path /path/to/data/images/dox

Option 2: Baseline Evaluation

Evaluate MLRMs on clean (unprotected) images.

bash run.sh --mode clean --step eval --dox_path /path/to/data/images/dox

Option 3: Generation Only

Generate adversarial images without evaluation.

bash run.sh --mode adv --step gen --dox_path /path/to/data/images/dox

🏋️ Training (Optional)

To train from scratch using GeoPrivacy-6K:

1. Ensure geo_6k images are in data/images/geo_6k.
2. Run the training script (initialization weights will be handled automatically):

bash scripts/train.sh

📜 Citation

@inproceedings{zhang2026disrupting,
  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},
  booktitle={International Conference on Learning Representations (ICLR)},
  year={2026}
}

📄 License

This project is licensed under the MIT License.