Detecting Discrete Cosine Transform-Based Digital Watermarking Insertion Area Using Deep Learning

Invisible digital watermarking, a technology for embedding information in digital content, is mainly used for copyright protection. In this paper, we proposed a method to identify images with invisible discrete cosine transform (DCT)-based watermarking and its vulnerability. Prior to watermarking, the images were normalized to 256 × 256 grayscale, and the network was generated through transfer learning; ResNet-18 was used to classify the input images as “watermarked” or “unwatermarked.” According to our results, the accuracy of the network, when classifying images into both classes, was as high as 99.80%. Furthermore, the testing accuracy of a network designed to detect the embedding location of the watermark in the frequency domain was 97.97%. It should be noted that the networks were ineffective when the DCT block size of the input images differed from that of the images in the training set.