摘要：Images available on online sharing platforms have a high probability of being modified, with additional global transformations such as compression, resizing or filtering covering the possible alteration. Such manipulations impose many constraints on forgery detection algorithms. This article presents a framework improving robustness for image forgery detection. The most important step of our framework is to take into account the image quality corresponding to the chosen application. Therefore, we relied on a camera identification model based on convolutional neural networks. Lossy compression such as JPEG being considered as the most common type of intentional or inadvertent concealment of image forgery, that leads us to experiment our proposal on this manipulation. Thus, our trained CNN is fed with a mixture of different qualities of compressed and uncompressed images. Experimental results showed the importance of this step to improve the effectiveness of our approach against recent literature approaches. To better interpret our trained CNN, we proposed an in-depth supervision by first a visualization of the layer and an experimental analysis of the influence of the learned features. This analysis led us to a more robust and accurate framework. Finally, we applied this improved system on an image forgery detection application and showed some promising results.
关键词：Forgery image detection ; Compression ; Camera model identification ; Convolutional neural networks supervision
