Detecting Double Compression in Images Using Difference of Class Activation Maps

In this paper a novel method for detecting double compression in images is proposed, which is a critical task in digital forensics. Grad-CAM (Gradient-weighted Class Activation Mapping) is utilized to analyze the gradients of a neural network and identify the differences between single- and doublecompressed images. The method is evaluated on a dataset of JPEG and WebP compressed images and its effectiveness in detecting double compression is demonstrated. The proposed method obtains good detection accuracy and AUC. A detailed analysis of the performance of the method on different compression levels and formats is provided, including a comparison of the detection accuracy and AUC for JPEG and WebP compression. The approach is found to be robust to various levels and formats of compression, making it a reliable tool for detecting double compression in images. Significant implications for digital forensics and image authentication are shown to exist, as the proposed method can be used to detect tampering and manipulation of images in a broader area.