摘要
公众舆论的系统性操纵，尤其是通过伪造图像在在线社交网络（OSN）中传播，已对社会构成严重威胁。本文提出一种名为数字取证网络（Digital Forensics Net, DF-Net）的深度神经网络，用于像素级图像伪造检测。所发布的模型在四个公认的基准数据集上均优于多项现有先进方法。尤为突出的是，DF-Net在面对社交网络中常见的有损图像操作（如图像缩放、压缩）时，仍具备出色的鲁棒性，这些操作通常由社交平台自动执行。
基准测试
| 基准 | 方法 | 指标 |
|---|---|---|
| image-manipulation-detection-on-casia-osn | DF-Net | AUC: 0.905 F-score: 0.587 Intersection over Union: 0.492 |
| image-manipulation-detection-on-casia-osn-1 | DF-Net | AUC: 0.902 Intersection over Union: 0.467 f-Score: 0.564 |
| image-manipulation-detection-on-casia-osn-2 | DF-Net | AUC: 0.905 Intersection over Union: 0.495 f-Score: 0.588 |
| image-manipulation-detection-on-casia-osn-3 | DF-Net | AUC: 0.902 Intersection over Union: 0.490 f-Score: 0.584 |
| image-manipulation-detection-on-casia-v1 | DF-Net | AUC: .906 |
| image-manipulation-detection-on-columbia | DF-Net | AUC: 0.880 |
| image-manipulation-detection-on-columbia-osn | DF-Net | AUC: 0.883 Intersection over Union: 0.681 f-Score: 0.760 |
| image-manipulation-detection-on-columbia-osn-1 | DF-Net | AUC: 0.881 Intersection over Union: 0.681 f-Score: 0.759 |
| image-manipulation-detection-on-columbia-osn-2 | DF-Net | AUC: 0.883 Intersection over Union: 0.685 f-Score: 0.762 |
| image-manipulation-detection-on-columbia-osn-3 | DF-Net | AUC: 0.890 Intersection over Union: 0.684 f-Score: 0.766 |
| image-manipulation-detection-on-dso-1 | DF-Net | AUC: 0.769 |
| image-manipulation-detection-on-dso-osn | DF-Net | AUC: 0.770 Intersection over Union: 0.245 f-Score: 0.359 |
| image-manipulation-detection-on-dso-osn-1 | DF-Net | AUC: 0.765 Intersection over Union: 0.245 f-Score: 0.358 |
| image-manipulation-detection-on-dso-osn-2 | DF-Net | AUC: 0.765 Intersection over Union: 0.249 f-Score: 0.361 |
| image-manipulation-detection-on-dso-osn-3 | DF-Net | AUC: 0.759 Intersection over Union: 0.245 f-Score: 0.354 |
| image-manipulation-detection-on-nist-osn | DF-Net | AUC: 0.795 Intersection over Union: 0.229 f-Score: 0.304 |
| image-manipulation-detection-on-nist-osn-1 | DF-Net | AUC: 0.799 Intersection over Union: 0.238 f-Score: 0.314 |
| image-manipulation-detection-on-nist-osn-2 | DF-Net | AUC: 0.803 Intersection over Union: 0.247 f-Score: 0.324 |
| image-manipulation-detection-on-nist-osn-3 | DF-Net | AUC: 0.791 Intersection over Union: 0.230 f-Score: 0.303 |