论文标题
具有可解释性的图像伪造检测
Image Forgery Detection with Interpretability
论文作者
论文摘要
在这项工作中，我们提出了一种基于学习的方法，重点是卷积神经网络（CNN）体系结构，以检测这些伪造。我们考虑检测基于复制的伪造和基于伪造的伪造的检测。为此，我们合成了自己的大数据集。除分类外，重点还放在伪造检测的解释性上。随着CNN分类产生图像级标签，重要的是要了解锻造区域是否确实有助于分类。为此，我们证明了使用Grad-CAM热图（在各种正确分类的示例中，伪造的区域确实是对分类的区域）的证明。有趣的是，这也适用于小型锻造区域，如我们的结果所示。这样的分析还可以帮助建立分类的可靠性。
In this work, we present a learning based method focusing on the convolutional neural network (CNN) architecture to detect these forgeries. We consider the detection of both copy-move forgeries and inpainting based forgeries. For these, we synthesize our own large dataset. In addition to classification, the focus is also on interpretability of the forgery detection. As the CNN classification yields the image-level label, it is important to understand if forged region has indeed contributed to the classification. For this purpose, we demonstrate using the Grad-CAM heatmap, that in various correctly classified examples, that the forged region is indeed the region contributing to the classification. Interestingly, this is also applicable for small forged regions, as is depicted in our results. Such an analysis can also help in establishing the reliability of the classification.
