The research on forgery detection and localization is significant in digital forensics and has attracted increasing attention recently. Traditional methods mostly use handcrafted or shallow-learning based features, but they have limited description ability and heavy computational costs. Recently, deep neural networks have shown to be capable of extracting complex statistical features from high-dimensional inputs and efficiently learning their hierarchical representations. In order to capture more discriminative features between tampered and non-tampered regions, we propose an improved mask regional convolutional neural network (Mask R-CNN) which attach a Sobel filter to the mask branch of Mask R-CNN in this paper. The Sobel filter acts as an auxiliary task to encourage predicted masks to have similar image gradients to the groundtruth mask. The overall network is capable of detecting two different types of image manipulations, including copy-move and splicing. The experimental results on two standard datasets show that the proposed model outperforms some state-of-the-art methods.

Exemplar-based image inpainting technology is a “double-edged sword”. It can not only restore the integrity of image by inpainting damaged or removed regions, but can also tamper with the image by using the pixels around the object region to fill in the gaps left by object removal. Through the research and analysis, it is found that the existing exemplar-based image inpainting forensics methods generally have the following disadvantages: the abnormal similar patches are time-consuming and inaccurate to search, have a high false alarm rate and a lack of robustness to multiple post-processing combined operations. In view of the above shortcomings, a detection method based on long short-term memory (LSTM)-convolutional neural network (CNN) for image object removal is proposed. In this method, CNN is used to search for abnormal similar patches. Because of CNN’s strong learning ability, it improves the speed and accuracy of the search. The LSTM network is used to eliminate the influence of false alarm patches on detection results and reduce the false alarm rate. A filtering module is designed to eliminate the attack of post-processing operation. Experimental results show that the method has a high accuracy, and can resist the attack of post-processing combination operations. It can achieve a better performance than the state-of-the-art approaches.

Image inpainting can effectively repair damaged areas, but it can also be a way of image tampering when it is used to remove meaningful content from an image. Therefore, this paper focuses on the research of inpainting forensics, and proposes a multi-task deep learning method. In order to enhance the learning of texture features, the corresponding local binary pattern channels are added to the input of the network. Furthermore, considering that the multi-task object detection network Mask R-CNN cannot fully utilize the features of all scale feature information during the FPN feature extraction phase, the network in this paper combines Feature Pyramid Networks and back connections to extract more features. This network model can detect not only the images tampered by traditional inpainting methods, but also the images inpainted by deep learning methods. Experimental results on two large public data sets demonstrate the superior performance of the proposed method.

Recently, generating visually secure cipher images by compressive sensing (CS) techniques has drawn much attention among researchers. However, most of these algorithms generate cipher images based on direct bit substitution and the underlying relationship between the hidden and modified data is not considered, which reduces the visual security of cipher images. In addition, performing CS on plain images directly is inefficient, and CS decryption quality is not high enough. Thus, we design a novel cryptosystem by introducing vector quantization (VQ) into CS-based encryption based on a 3D fractional Lorenz chaotic system. In our work, CS compresses only the sparser error matrix generated from the plain and VQ images in the secret generation phase, which improves CS compression performance and the quality of decrypted images. In addition, a smooth function is used in the embedding phase to find the underlying relationship and determine relatively suitable modifiable values for the carrier image. All the secret streams are produced by updating the initial values and control parameters from the fractional chaotic system, and then utilized in CS, diffusion, and embedding. Simulation results demonstrate the effectiveness of the proposed method.