An image authentication and recovery system based on discrete wavelet transform and convolutional neural networks

In recent years, research on image authentication and recovery using convolutional neural networks (CNNs) has attracted tremendous attention. Most existing studies on this topic treat such authentication and recovery simply as a type of pixel-wise annotation and prediction, that is, annotation of true/false labels on each pixel and predicting the value of each tampered pixel. However, such machine learning mechanisms show unstable performance and effectiveness. Classical digital watermarking techniques can secure the integrity of the data; thus, they can leverage the data to stabilize the machine-learning-based image authentication and recovery performance . Accordingly, this paper proposes a model that combines CNNs and classical digital watermarking techniques (DWTs); this assures the integrity of data by using the traditional image authentication method in the DWT domain. We examined the effectiveness and performance of the proposed approach against a few of the most popular image tampering attacks. We also compared our approach with several state-of-the-art works. Experimental results show that our proposal provides predominant tampering recognition. In addition, our proposed method can precisely recover tampered regions of a image.