Digital Forensics and Watermarking

With the increasing need for secure image transmission and storage, researchers have focused on developing advanced encryption algorithms. This article highlights the significance of a newly constructed 1D chaotic system and its application in constructing a secure image encryption algorithm over blockchain. The newly proposed 1D logistic cosine tangent chaotic system consists of the cosine function and tangent function replacing the two variables in the logistic map, respectively, thus improving the dynamics of the new 1D chaotic map. The image encryption scheme constructed based on the newly proposed 1d chaotic system relies on the privilege management function of the blockchain’s smart contract to design the key transfer scheme, which simultaneously achieves the security of the cryptographic algorithm and the reliability of the key transfer. The experimental results provide strong evidence that the proposed cryptographic system ensures robust security. This research provides a novel and effective solution for ensuring the confidentiality and integrity of encrypted images, addressing the growing concerns of image security in the digital era.

The utilization of Photo Response Non-Uniformity (PRNU) technology has found extensive application in the field of multimedia forensics, particularly in the authentication of the original camera source of an image. However, this technique has also given rise to significant concerns regarding privacy breaches. For instance, adversaries can exploit publicly available images to generate PRNU and subsequently impersonate the owners of the images. In response to these challenges, we propose an algorithm for achieving source device anonymity in widely used JPEG images. The method combines the discrete cosine transform (DCT) with JPEG compression to process the DCT coefficients of an image after inverse quantization. By ensuring the high quality of the processed image, this approach effectively breaks the link between an image and its source camera. Additionally, a reversible data hiding method is employed, enabling the recovery of traceability if necessary. Our algorithm offers several advantages over existing schemes. It operates within the domain of JPEG image compression, maintaining a low time complexity. Additionally, it effectively preserves the visual quality of images and eliminates the typical traceability effects associated with images.

Intrinsic, non-invasive product authentication is the preferred way of detecting counterfeit products as it does not generate additional costs during the production process. Previous works achieved promising results for smartphone-based product authentication. However, while promising, the methods fail when enrollment and authentication are performed on different devices (cross-device). This work investigates the underlying reasons for the limitations in the practical application of cross-device intrinsic surface structure-based product authentication. In particular by utilising micro-texture classification approaches applied on images of zircon oxide blocks (dental implants) captured using a commodity smartphone device. The main result is that the device-specific artefacts (image sensor as well as image processing-specific ones) are so strong that they obfuscate the material microstructure. To be more precise, the device’s intrinsic signal makes device identification easier to perform than the material authentication.

As the use and reliance on AI technologies continue to proliferate, there is mounting concern regarding adversarial example attacks, emphasizing the pressing necessity for robust defense strategies to protect AI systems from malicious input manipulation. In this paper, we introduce a computationally efficient plug-in module, seamlessly integrable with advanced diffusion models for purifying adversarial examples. Drawing inspiration from the concept of deconstruction and reconstruction (DR), our module decomposes an input image into foundational visual features expected to exhibit robustness against adversarial perturbations and subsequently rebuilds the image using an image-to-image transformation neural network. Through the collaborative integration of the module with an advanced diffusion model, this combination attains state-of-the-art performance in effectively purifying adversarial examples while preserving high classification accuracy on clean image samples. The model performance is evaluated on representative neural network classifiers pre-trained and fine-tuned on large-scale datasets. An ablation study analyses the impact of the proposed plug-in module on enhancing the effectiveness of diffusion-based purification. Furthermore, it is noteworthy that the module demonstrates significant computational efficiency, incurring only minimal computational overhead during the purification process.

With the advancement of cloud computing, outsourced image processing has become an attractive business model, but it also poses serious privacy risks. Existing privacy-preserving image scaling techniques often use secret sharing schemes or the Paillier cryptosystem to protect privacy. These methods require the collaboration of multiple servers or only support additive operations, which increases the difficulty of data storage and complicates image processing. To address these issues, this paper focuses on cloud-based privacy-preserving image scaling in the encrypted domain. We propose an image scaling scheme based on homomorphic encryption, which allows cloud servers to perform scaling operations on encrypted images using bicubic interpolation. To avoid the high storage and communication costs of per-pixel encryption, we introduce an efficient data encoding method where a single ciphertext contains the information of an entire image. This significantly reduces the storage space and communication overhead with the cloud server, while our scheme supports computational operations in this data format. Our experimental results validate the feasibility of the proposed scheme, which outperforms existing schemes in terms of storage overhead and operational efficiency.

With the development of communication technologies, the practice of creating new texts by manipulating original sentence structures through multi-turn machine translation is widespread across various domains. Existing plagiarism detection models often treat different features uniformly and overlook the significance of disparities within high-dimensional features. Therefore, this paper proposes a novel plagiarism detection model towards multi-turn text back-translation (PDMTT), adopting a novel mechanism that combines local and global features and enhances them. The grouping enhancement fusion (GEF) mechanism assigns importance coefficients to sub-features, reinforcing critical aspects while diminishing less relevant ones. These enhanced features, generated by the GEF mechanism, are leveraged to extract high-quality text representations, thereby improving the precision of the model in distinguishing original content from back-translated texts. Furthermore, we improve the back-translation plagiarism detection capability of our model by optimizing the contrastive loss function and utilizing the fused translated representations as targets. To validate the effectiveness of our model, we also constructed a multi-tuple back-translation plagiarism dataset for model training and validation. Experimental results demonstrate that the proposed PDMTT outperforms previous methods in back-translation plagiarism detection, yielding superior text representations. The ablation study further confirms that the incorporation of the GEF mechanism effectively enhances the discrimination capability of our model.

The conventional perceptual hashing algorithms are constrained to a singular global feature extraction algorithm and lack efficient scalability adaptation. To address this problem, an image-perceptual hashing algorithm based on convolutional neural networks is proposed in this paper. First of all, the entire image is convolved by the backbone network to obtain a feature map. The Region Proposal Network (RPN) is employed to generate multiple-sized proposal frames at each location by using sliding windows. Considering the complexity and diversity of the object, proposal boxes of various sizes and shapes are formulated, and the local features are comprehensively exploited in an image, thereby, generating a perceptual hash code that can represent the semantic features of an image strongly. Moreover, The Mean Square Error (MSE) loss is incorporated into the optimization process to evaluate the coincidence between the proposal frame and the actual frame, generating more representative hash codes. Finally, an image perceptual hash code with high intuitive features can be formulated through iterative training of the proposed convolutional neural networks. Extensive experimental results demonstrate that the proposed image perceptual hashing algorithm based on a convolutional neural network surpasses other state-of-the-art methods.

GAN-based I2I translation techniques for unpaired data are employed for the synthesis of biometric finger vein presentation attack instrument samples corresponding to three public presentation attack datasets. For the assessment of these synthetic samples, we analyse their behaviour when attacking finger vein recognition systems, comparing these results to such obtained from actually crafted presentation attack samples. We observe that although visual appearance and sample set correspondence are surprisingly good for some networks, respectively, the assessment of the behaviour of the data in a conducted attack is more difficult. Even if for some recognition schemes out of 11 considered we find a good accordance in terms of IAPMR (for many we don’t), the attack score distributions turn out to be highly dissimilar when comparing crafted and synthetic presentation attack instrument samples. More work is needed to be able to correctly interpret corresponding diverging results with respect to the relevance in attack simulation. From the seven network architectures considered, CycleGAN provides the most useful results, but the artificially created samples do not fully mimic the behaviour of crafted ones.

Deep video inpainting can automatically fill in missing content both in spatial and temporal domain. Unfortunately, malicious video inpainting operations can distort media content, making it challenging for viewers to detect inpainting traces due to their realistic visual effects. As a result, the detection of video inpainting has emerged as a crucial research area in video forensics. Several detection models that have been proposed are trained and tested on datasets made by three kinds of inpainting models, but not tested against the latest and better deep inpainting models. To address this, we introduce a novel end-to-end video inpainting detection network, comprising a feature extraction module and a feature learning module. The Feature extraction module is a Bayar layer and the feature learning module is an encoder-decoder module. The proposed approach is evaluated with inpainted videos created by several state-of-the-art deep video inpainting networks. Extensive experiments has proven that our approach achieved better inpainting localization performance than other methods.

To counter the counterfeit of 3D printed products, 3DPS, an improved printing signature for 3D printed objects based on our previous work is proposed for authentication. A specific hole is added to the non-critical flat portion of the 3D printed object, and the 3DPS is constructed from the contour of the hole with a hand-held microscope. Compared to the previous work, the equipment distortion model and the 3DPS’s construction are improved, and the threshold can be calculated directly rather than determined by experience. Experimental results show that the 3DPS not only can effectively authenticate 3D printed objects with high accuracy but is also robust and secure.

The advancement of Facial Attribute Editing (FAE) technology allows individuals to effortlessly alter facial attributes in images without discernible visual artifacts. Given the pivotal role facial features play in identity recognition, the misuse of these manipulated images raises significant security concerns, particularly around identity forgery. While existing image forensics algorithms primarily concentrate on traditional tampering methods like splicing and copy-move and are often tailored to detect tampering in natural landscape images, they fall short in pinpointing FAE manipulations effectively. In this paper, we introduce two FAE datasets and propose the Multi-Scale Enhanced Dual-Stream Network (MSDS-Net) specifically for FAE Localization. Our analysis reveals that FAE artifacts are present in both the spatial and DCT frequency domains. Uniquely, in contrast to traditional tampering methods where modifications are localized, facial attribute alterations often span the entire image. The transitions between edited and unedited regions appear seamless, devoid of any conspicuous local tampering signs. Thus, our proposed method adopts a dual-stream structure, targeting the extraction of tampering signs from both the spatial and DCT frequency domains. Within each stream, multi-scale units are employed to discern editing artifacts across varying receptive field sizes. Comprehensive comparative results indicate that our approach outperforms existing methods in the field of FAE localization, setting a new benchmark in performance. Additionally, when applied to the task of pinpointing facial image inpainting, our method demonstrated commendable results.

Neural networks have been applied in various fields, including steganography (called neural network steganography). The network used for secret data extraction is called the extractor. This paper proposes a neural network steganography scheme using extractor matching. In our scheme, the extractor is a publicly available normal network possessed by the receiver, which is used for conventional intelligent tasks. Sender connects extractor to another neural network (called cover network), and then trains the connected network to guarantee correctly data extraction without decreasing the performance of the original task of cover network. During the process of training, the parameters of extractor remain unchanged. Specifically, these network parameters are obtained using an extraction key. The receiver can correctly extract secret data with the help of correct extraction key, while an incorrect key will fail to extract secret data. The feasibility of our scheme is demonstrated in experiments.

Recent studies have demonstrated that generative models, such as Generative Adversarial Networks (GANs), leave discernible traces in their results. Based on these traces, several forensic methods have achieved remarkable detection accuracy and strong generalization across different generative models. To counter forensic methods and identify potential vulnerabilities in detectors, existing anti-forensics methods primarily focus on embedding adversarial noises into spacial images. In addition, most methods design distinct noise patterns to each image, making it challenging to generate many adversarial samples within a short time. To address these limitations, this paper proposes a novel anti-forensics method in the frequency domain via using image pairs generated with GAN inversion technology. The objective is to design a universally effective approach that avoids introducing noticeable spatial traces. The proposed method introduces a fresh perspective by applying GAN inversion technology to the field of frequency-domain anti-forensics and only requires 100 images, which is effective to handle all the outputs of the target generator and to generate numerous adversarial samples in turn to help enhance the performance of the detector. Our experiment results show a significant reduction of the detection performance. Specially, when two target models detect both generated and edited images based on the StyleGAN, the area under the receiver-operating curve (AUC) decreases by 9.0\(\%\).

Traditional steganographic algorithms often suffer from issues such as low visual quality and limited resilience against steganalysis at high-capacity data embedding. To address these limitations, this paper proposes a cross-channel image steganography algorithm based on generative adversarial networks. In contrast to conventional image steganography techniques that directly embed secret data into original carrier images, the proposed algorithm embeds the secret data into the difference-plane of the two similar color channels. The proposed data embedding scheme involves a U-Net structure based generator for steganography, an adversarial network for steganalysis, and an optimization network for enhancing anti-steganalysis capabilities. In addition, a newly introduced Lion optimizer is introduced to effectively optimize the convergence speed of the proposed networks by adaptively setting learning rates and weight decay values. At the same time, the mean square error loss, structural similarity loss, and adversarial loss are employed to progressively enhance the visual quality of generated stego images. Consequently, a color image can be seamlessly embedded into the same-sized color image, and achieving high perceptual quality. Experimental results demonstrate that the proposed algorithm achieves a peak PSNR of 41.6 dB for color stego images, significantly reducing the distortion caused by secret image embedding.

With the rapid advancement of cloud computing, great security concern regarding data storage has been brougth forth. This paper proposes a reversible data hiding algorithm for JPEG images based on Paillier homomorphic encryption. The secret data is embedded by modifying quantized DCT coefficients. The AC coefficients are modified through homomorphic encryption during the embedding process, while the DC coefficients, which significantly impact image quality, are subjected only to Arnold transform. Additionally, an algorithm that can control the distribution range of ciphertext of Paillier homomorphic encryption is frist presented, thereby, minimizing the ciphertext expansion. Extensive experimental results demonstrate that the proposed algorithm achieves larger data embedding capacity, higher security of ciphertext images, and more stable image quality in comparison with other advanced algorithms.

Reversible data hiding techniques can effectively solve the information security problem, and One crucial approach to enhance the level of reversible data hiding is to predict images with higher accuracy, thereby effectively increasing the embedding capacity. However, conventional prediction methods encounter limitations in fully exploiting global pixel correlation. In this study, we propose a novel framework that combines image splitting and convolutional neural network (CNN) techniques. Specifically, grayscale images are divided into non-overlapping blocks and categorized into texture and smoothing groups based on mean square error calculations for each block. This strategy not only improves operational efficiency but also enhances prediction accuracy. Additionally, we leverage the multi-sensory field and global optimization capability of CNNs for image prediction. By assigning image blocks to specific predictors targeting either texture or smoothing groups according to their respective categories, more precise predicted images can be obtained. The image predictor is trained using 3000 randomly selected images from ImageNet. The experimental results show that the proposed method can predict images accurately and improve the prediction performance more effectively. Compared with other methods, the overall performance of the method is higher.

Automatic polyp segmentation is a crucial application of artificial intelligence in the medical field. However, this task is challenging due to uneven brightness, variable colors, and blurry boundaries. Most current polyp segmentation methods focus on features extracted from the spatial domain, ignoring the valuable information contained in the frequency domain. In this paper, we propose a Dual-Domain Learning Network (D\(^{2}\)LNet) for polyp segmentation. Specifically, we propose a Phase-Amplitude Attention Module, which enhances the details in the phase spectrum, while reducing interference from brightness and color in the amplitude spectrum. Moreover, we introduce a Spatial-Frequency Fusion Module that utilizes parameterized frequency-domain features to adjust the style of spatial-domain features and improve polyp visibility. Extensive experiments demonstrate that our method outperforms the state-of-the-art approaches both visually and quantitatively.

In the contemporary era of the digital economy, the significance of data as a fundamental production factor cannot be overstated, as it serves as a driving force behind the development of the digital economy. Despite the existence of numerous algorithms that have been proposed to achieve intersection of datasets while ensuring privacy, challenges related to computational efficiency and trust costs between institutions persist. This paper presents the BTPSI (Blockchain-based Two-round Private Set Intersection) algorithm, which aims to efficiently and reliably compute the intersection of datasets within the unbalanced private set intersection framework. The proposed BTPSI algorithm has been thoroughly tested and evaluated on various datasets of varying sizes.

This paper provides an in-depth discussion on the application of quaternion orthogonal Fourier-Mellin Moments (QOFMM) in the field of digital image processing, and proposes a novel method of factorial operation aiming to optimize its computational efficiency and accuracy. In addition, this paper focuses on the importance of zero-watermark technology in the field of information security. QOFMM is an advanced feature mention technique, which is particularly applicable to the field of digital image processing and information security. In this technique, the factorial operation plays a key role. However, traditional factorial computation methods may encounter efficiency bottlenecks when dealing with large data, thus affecting the overall performance. To address this issue, this study proposes an innovative method for factorial operation, which performs factorial operation by improving the radial basis function computation strategy, aiming to reduce the computational complexity and enhance the computational accuracy. To verify the effectiveness of the proposed method, we compare it with existing methods of factorial computation. The experimental results show that the new method significantly improves both processing speed and computational capability. Overall, this paper provides a new QOFMM optimization strategy, which not only improves the computational efficiency and accuracy but also brings a new research direction to the field of digital image processing and zero-watermarking technology.

Steganalysis, as an adversarial technique to steganography, aims to uncover potential concealed information transmission, holding significant research implications and value in maintaining societal peace and stability. With the rapid development and application of DNA synthesis technology, an increasing number of information hiding technologies based on DNA synthesis have emerged in recent years. DNA, as a natural information carrier, boasts advantages such as high information density, robustness, and strong imperceptibility, making it a challenging target for existing steganalysis technologies to efficiently detect. This paper proposes a DNA steganalysis technique that integrates multi-dimensional features. It extracts short-distance and long-distance related features from the DNA long chain separately and then employs ensemble learning for feature fusion and discrimination. Experiments have shown that this method can effectively enhance the detection capability against the latest DNA steganography technologies. We hope that this work will contribute to inspiring more research on DNA-oriented steganography and steganalysis technologies in the future.

In recent years, more and more steganographic methods based on streaming voice have appeared, which poses a great threat to the security of cyberspace. In this paper, in order to promote the development of streaming voice steganalysis technology, we construct and release a large-scale streaming voice steganalysis dataset called VStego800K. To truly reflect the needs of reality, we mainly follow three considerations when constructing the VStego800K dataset: large-scale, real-time, and diversity. The large-scale dataset allows researchers to fully explore the statistical distribution differences of streaming signals caused by steganography. Therefore, the proposed VStego800K dataset contains 814,592 streaming voice fragments. Among them, 764,592 samples (382,296 cover-stego pairs) are divided as the training set and the remaining 50,000 as testing set. The duration of all samples in the data set is uniformly cut to 1 s to encourage researchers to develop near real-time speech steganalysis algorithms. To ensure the diversity of the dataset, the collected voice signals are mixed with male and female as well as Chinese and English from different speakers. For each steganographic sample in VStego800K, we randomly use two typical streaming voice steganography algorithms, and randomly embed random bit with embedding rates of 10%–40%. We tested the performance of some latest steganalysis algorithms on VStego800K, with specific results and analysis details in the experimental part. We hope that the VStego800K dataset will further promote the development of universal voice steganalysis technology. The description of VStego800K and instructions will be released here: https://​github.​com/​YangzlTHU/​VStego800K.

With the rapid development of the Internet, more and more methods of text steganography have emerged. However, these methods are easily abused in public networks for malicious purposes, which poses a great threat to cyberspace security. At present, a large number of text steganalysis methods have been proposed to game with text steganography. However, existing methods typically assume a balanced class distribution. In reality, stego texts are far less than cover texts. How to accurately detect stego texts in massive texts becomes a challenge. In this paper, we propose a text steganalysis method based on an under-sample method and ensemble learning in imbalanced scenarios. Specifically, we introduce the thinking of clustering to under-sample the majority class samples (cover texts) based on the detection difficulty of the samples, in order to select samples with rich information. Ensemble learning is then used to ensemble the detection results of multiple base classifiers and guide the sampling process. We designed several experiments to test the detection performance of the proposed model. Experimental results show that the proposed model can effectively compensate for the deficiencies of existing methods, even in highly imbalanced datasets, the model can still detect stego texts effectively.