ABSTRACT Deep learning methods for image forgery detection often struggle with compression attack robustness. This paper proposes a novel multi‐class forgery detection framework combining spatial‐frequency fusion with Swin‐Transformer, outperforming existing methods in compression attack scenarios. Our approach integrates a frequency domain perception module with quantization tables, a spatial domain perception module through multi‐strategy convolutions, and a dual‐attention mechanism combining spatial and channel attention for feature fusion. Experimental results demonstrate superior performance with an F 1 score of 87% under JPEG compression ( q = 75), significantly surpassing current state‐of‐the‐art methods by an average of 15% in compression resistance while maintaining high detection accuracy.

The commonly used standard convolutional layers cannot adaptively adjust the number and locations of sampling points according to the scales and shapes of tampered regions, which increases the difficulty of detecting images containing tampered regions of different sizes. Therefore, the selective sampling attention (SSA) is proposed to automatically learn the number and locations of sampling points as well as the weight of each sampling point within a certain context range of the input feature map through backpropagation, which can help the network better adapt to tampered regions of different scales and shapes. In addition, the self-correlation calculation (SCC), aiming at calculating the similarity between every two feature points in a feature map, necessarily incurs an expensive computational burden when used for high-resolution feature maps. To remedy the problem, the two-step SCC (TS-SCC) with low computation burden is proposed to pick out highly similar regions by means of the feature similarity obtained from low-resolution version of the input feature map, so that the high-resolution version merely needs to calculate the similarity between every two feature points within its high-similarity regions. Finally, to predict the edges and interiors of copy-move tampered regions more precisely, adaptive dual-branch feature fusion module is proposed to employ a lightweight multi-scale atrous convolutional module to adaptively fuse multi-level features before TS-SCC and the correlation features after TS-SCC, thereby improving the detection performance. Combining these three structures, a lightweight, fast, low-cost and high-precision CMFD network, ST-Net, is designed in this paper. Experimental results on four publicly available datasets verify that ST-Net outperforms several related CMFD networks in terms of detection accuracy, number of parameters, computational cost and inference time.

Existing convolutional neural network-based reversible data hiding (RDH) predictors typically stack the standard convolution blocks with stride 1 for feature extraction, and keep the sizes of input and output feature maps unchanged through padding. This suggests that only a limited range of contextual spatial information is obtained. To remedy this problem above, a U-Net-like RDH predictor named UCANet is proposed in this paper to capture rich multi-scale contextual information by gradually downsampling feature maps. To fuse two feature maps at different levels along the channel dimension, we put forward the channel adaptive attention (CAA). By merely combining cheap pointwise convolution operations, CAA achieves the integration of non-linear and linear features as well as implicitly enhances channel dimensionality with low computational burden, thereby effectively enriching the expression of the channel information. The design of UCANet considers the characteristics of RDH from two aspects. On the one hand, instead of maxpooling or average pooling commonly used for downsampling, a stride-2 convolution block that can adaptively adjust the weights of convolution kernels and select useful information is utilized to downsample feature maps. On the other hand, UCANet removes the batch normalization layers to avoid their influence on the distribution of feature maps, which helps to strengthen the network's prediction capability. Extensive experiments also demonstrate that the proposed UCANet achieves better prediction performance, compared to several state-of-the-art methods.

The existing deep learning based copy-move forgery detection (DL-CMFD) networks focus on providing impressive detection accuracy for tampered regions of different sizes, but usually result in high computation cost and a large number of parameters. The focus of this letter is to propose LHCM-Net, a lightweight, high-precision DL-CMFD network, by integrating a low-cost self-correlation calculation (SCC) module (LCSCC), a gated feature fusion module (GFFM) and residual U-blocks (RSU) equipped with FasterNet blocks (FRSU). Considering that SCC, which calculates the similarity between every two pixels, inevitably leads to high computation cost, existing DL-CMFD networks have to carry out SCC only on low-resolution feature maps to reduce the computation cost. To make high-resolution features available for SCC without obviously introducing high computation cost, this letter proposes LCSCC to calculate the similarity between pixels with a certain distance. GFFM is presented to fuse feature maps of different spatial resolutions by adaptively adjusting their weights based on their respective characteristics, thereby fully integrating high-resolution and low-resolution features for subsequent LCSCC and obviously enhancing the detection accuracy. The FRSU allows LHCM-Net to keep the number of parameters (NP) and computation cost low by combining lightweight FasterNet blocks. The experimental results also demonstrate that LHCM-Net outperforms several existing DL-CMFD networks on three publicly available datasets in terms of detection accuracy, NP and computation cost.

Abstract In the metaverse era, virtual design of hairstyle becomes very popular for personalized aesthetics. As hair design tasks can be decomposed into hair attribute editing and generation, the development of generative adversarial networks (GANs) has significantly prompted its development. The majority of the existing algorithms focus on transferring the overall hair region from one face to another, which ignore fine control over the color and geometric features. Furthermore, these algorithms may result in unnatural generation results. In this paper, we propose a hair modification framework that learns hairstyle information from a reference face mask and color information from a guidance face image. Firstly, the features of the input face image and reference images are extracted through a group of encoders, and then divided into feature vectors of coarse, medium, and fine levels. Secondly, multi-level feature vectors are fused in the latent space using attention-based modulation modules. Finally, the fused feature vector is passed through a StyleGAN generator to generate face images with specified hairstyle and hair color. Experimental results show that the proposed method can finely simulate the hairstyle transition between long and short hair under the constraint of the reference mask, and can produce realistic fusion effects in the hair-covered regions, such as ears, neck, and forehead. Various hair dyeing effects that adapt to personalized characteristics are demonstrated, as facial features including skin color and hair texture are preserved when transferring the hair color.

Model watermarking Model watermarking is a method for embedding watermark information into a neural network model. It proves the ownership of the model without affecting its performance. Since there are plenty of attacks against model pruning, it becomes more significant to design anti-pruning model watermarking algorithms. In this paper, multiple watermark embedding is performed to protect the model copyright for the image steganography auto-encoder model “Hiding Data with Deep Networks” (HiDDeN). Firstly, the appropriate model weights are selected by employing three classical model pruning algorithms of model weights. Secondly, the model watermark is spread by using Discrete Cosine Transform (DCT)-based image watermarking algorithm, which improves the noise and pruning resistance of the model watermark. Finally, the model watermark is embedded to the 4th and 5th decimal places of the selected model weights. The experimental results demonstrate that the proposed algorithm has a good robustness against model pruning without affecting the watermark extraction performance of the auto-encoder network model. Even with the embedded model watermark, the decoder's watermark extraction accuracy is still higher than 0.9993. and the autoencoder is still valuable when the model weights are pruned by 40%. Furthermore, the proposed algorithm has a certain degree of improvements in watermarking capacity.

We present stego-marbling-texture, a new and unique texture design method which allows users to deliver personalized messages with beautiful marbling textures. Our approach is inspired by the success of the recent work on modeling traditional marbling operations as mathematical functions. The encrypter transforms an input image or a text message into an intricate marbling pattern using marbling operations defined as reversible functions, and the decrypter recovers the input image or message through reversing the process of marbling operations. When applying marbling operations, the parameters of operations are automatically recorded, encrypted, and then invisibly embedded into the marbling pattern to create a stego-marbling-texture. In this way, the decrypter can be implemented as a stand along software, enabling the receiver to extract the hidden message from the stego-marbling-texture without requiring any extra information from the sender. To ensure that the message is unnoticeably and beautifully covered by the marbling texture, we propose a new technique for automatically creating a background which is harmonious with the input message based on a set of visual perception cues.

This paper presents a public mesh watermarking algorithm. First, the resultant watermarked image minus the original image is the watermark information. According to the addition property of the Fourier transform, a change of spatial domain will cause a change in the frequency domain. Then, the watermark information is scaled down and embedded in one part of the x-coordinate of the original mesh. Finally, the x-coordinate of the test mesh is amplified before extraction. Experimental results prove that our algorithm is resistant to a variety of attacks without the need for any preprocessing.

For the classical watermark embedment model I = 1 + /spl alpha/W, the corresponding watermark detection has its limitation in its need of a fixed parameter for extracting watermarks. If the extraction parameter is too large, we cannot extract the watermark from the image that contains watermarks; if it is too small, the extracted watermarks may be blurred. This paper proposes a novel watermark extraction method. First, we treat the watermark information as noise for the watermarked image in its spatial domain. We then magnify the noise before detection. Next, we recover the watermark information by adjusting the extracted data from the frequency domain according to our global minimum method. Experimental results show that our watermark extraction method is more valid and accurate than the classical method. It can greatly reduce extraction errors.