Overcoming Natural Similarity: A DenseNet-Based Approach for Reliable Transform Copy-Move Forgery Identification

초록

Digital image forensics aims to verify the authenticity of an image by determining its origin and any processing it has undergone. Among passive approaches in copy move forgery detection, which analyze images without requiring prior embedded information, a region of interest in an image is duplicated and manipulated to conceal or alter content. This process often involves multiple transformations that significantly alter the image. These transformations consist of several operations, such as geometrical transformations, frequency domain transformations, noise addition, etc. In copy-move forgery detection, conventional approaches are generally effective at identifying simple copy-move forgeries. However, these approaches often fall short in detecting copy-move forgeries when frequency domain transformations are applied to the copied regions. Further, existing approaches have a high false positive rate in terms of detecting natural similarity. The proposed method introduces an advanced deep learning approach for detecting copy-move forgeries in frequency domains. During preprocessing, segmentation is utilized to extract the objects or regions of interest in the image. DenseNet121 is employed to extract high-level feature from these segmented regions with a detailed analysis of features from each block. By comparing the features extracted from block 4, the method identifies and maps regions with similar features, thereby detecting potential transformed copy-move manipulations. This proposed method demonstrates robustness in identifying forgeries that involve transformations and significantly enhances detection accuracy with natural similarity. However, the method shows reduced performance on highly textured images, where natural similarity and repetitive patterns make discrimination more difficult.

키워드