No abstract

Nowadays, Image Forgery Detection contributes an indispensable role in digital forensics, while there are increasingly more sophisticated forgery methods. In overall, almost conventional methods just focus on identifying specific features in tampered images, therefore, such methods cannot cover whole possible cases in reality. Recently, some data-driven proposals have been exploited to handle these barriers and attained prominent results. However, almost these ones are hungry to data because of the complication in deep architectures, which requires a large amount of data and an energetic implementation hardware. In this paper, we propose a low computational-cost and effective data-driven model as a modified deep learning-based model to solve the existing problems above. The process of approach is overviewed as follows: Firstly, the Daubechies Wavelet transform is utilized to extract features of size 450, representing YCrCb patches inside the image. Then, a neural network is used to classify forged patches. However, when conducting a discrimination analysis, we found that the luminance channel (Y) does not play an essential role in the forgery detection, whereas, it is better by using two chrominance channels (Cr and Cb). The idea is stated by removing these luminance features, then the feature vector dimension changes to as two-thirds as its origin, which reduces efficiently the computational cost in both of training and testing processes. The experimental results reveal that our proposed method reaches a high detection accuracy of 97.11%, even the model suffers in some difficult circumstances (e.g., narrowness, and lack of positive training samples). As a result, the proposed model is effective to address the mentioned challenges.