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.

翻译：在这项工作中,我们展示了一种学习方法,侧重于进化神经网络(CNN)结构,以探测这些伪造物。我们考虑探测复制移动的伪造物和涂漆的伪造物。我们综合了自己的大数据集。除了分类外,我们还侧重于伪造检测的可解释性。随着CNN分类生成图像级标签,了解伪造区域是否确实有助于分类非常重要。为此,我们使用格拉德-CAM热映(Grad-CAM热映)来证明,在各种正确分类的例子中,伪造区域确实是对分类的贡献区域。有趣的是,这也适用于小型伪造区域,正如我们的结果所描述的那样。这种分析也有助于确定分类的可靠性。