---
license: apache-2.0
pipeline_tag: object-detection
library_name: transformers
---

# Adversarial Attention Perturbations for Large Object Detection Transformers (AFOG)

This repository hosts the implementation for the paper [Adversarial Attention Perturbations for Large Object Detection Transformers](https://huggingface.co/papers/2508.02987).

**AFOG** (Attention-Focused Offensive Gradient) is a novel adversarial perturbation method designed to attack both large transformer-based object detectors and conventional CNN-based detectors with a unified adversarial attention framework. It leverages a learnable attention mechanism to focus perturbations on vulnerable image regions, significantly increasing attack performance while maintaining visual imperceptibility.

<div align="center">
  <img src="https://github.com/IDEA-Research/AFOG/raw/main/assets/examples.png" width="100%" height="100%"/>
</div>

## Key Features
*   **Unified Attack Framework**: Effective against both transformer-based and CNN-based object detectors.
*   **Attention-Focused Perturbations**: Utilizes a learnable attention mechanism to target vulnerable image regions, improving attack efficacy.
*   **Efficient and Stealthy**: Generates strategically designed, visually imperceptible perturbations that cause well-trained models to fail.

## Installation and Usage
For detailed installation instructions, environment setup, and comprehensive usage examples (including Jupyter notebooks for single-image attacks and Slurm scripts for COCO evaluation), please refer to the official [AFOG GitHub repository](https://github.com/IDEA-Research/detrex).