DDA-COCO

# DDA-COCO Benchmark

### Official Dataset for **Dual Data Alignment Makes AI-Generated Image Detector Easier Generalizable**
**Conference:** 39th Conference on Neural Information Processing Systems (NeurIPS 2025) https://arxiv.org/abs/2505.14359

---

#### Dataset Description
**DDA-COCO** is a benchmark specifically designed to evaluate whether AIGI detectors rely on **"non-causal features"** (such as compression artifacts or content semantics).

Many existing detectors experience significant performance drops when tested on strictly aligned data, as they tend to learn dataset biases rather than intrinsic generation artifacts. DDA-COCO includes real images from the MSCOCO validation set and their corresponding synthetic images, processed with various VAE reconstructions and frequency alignments, to test detector robustness.

#### Content
The dataset contains 5 subsets corresponding to different VAE model reconstructions:
* **Source:** MSCOCO Validation Set (Real).
* **Variations:** Synthetic images reconstructed by different VAE versions (e.g., SD1.5, SD2.1, SDXL) with frequency alignment.
* **Key Feature:** High consistency between real and synthetic images in semantics, size, and frequency distribution, forcing detectors to focus on subtle generative traces.

#### Citation
```code
@inproceedings{chen2025dual,
title={Dual Data Alignment Makes {AI}-Generated Image Detector Easier Generalizable},
author={Ruoxin Chen and Junwei Xi and Zhiyuan Yan and Ke-Yue Zhang and Shuang Wu and Jingyi Xie and Xu Chen and Lei Xu and Isabel Guan and Taiping Yao and Shouhong Ding},
booktitle={The Thirty-ninth Annual Conference on Neural Information Processing Systems},
year={2025},
url={https://openreview.net/forum?id=C39ShJwtD5}
}
```

DDA-COCO是一款基于MSCOCO 2017验证集构建的高难度AI生成图像（Artificially Generated Imagery, AIGI）检测基准数据集。真实图像与经变分自编码器（Variational Autoencoder, VAE）重构且内容对齐的合成图像一一配对，所有图像均经过JPEG-96压缩以消除格式偏差，以此促使检测器聚焦于真正的生成伪影，而非无关紧要的压缩痕迹。