In-generation watermarking has emerged as a promising solution for provenance tracing and copyright protection of AI-generated images. Existing watermarking methods for Latent Diffusion Models (LDMs), however, still face a critical challenge in balancing generative fidelity and robustness against severe attacks. In this paper, we propose MsLW, a unified watermarking framework for LDMs that propagates watermark signals throughout the late-stage denoising trajectory, enabling the watermark to become an intrinsic component of the image generation process. Specifically, we introduce a lightweight Latent Watermarking Network (LWN) that learns the generation of watermark residual in the denoised latent space, together with a Multi-step Watermark Embedding (MsWE) strategy for progressively injecting the learned identical watermark residual into not only the denoised latent variable but also the intermediate latent ones. To mitigate distribution disturbance caused by iterative watermark propagation, MsWE further incorporates a latent-space distribution alignment mechanism, as well as wavelet low-pass filtering to suppress perceptual artifacts. Extensive experiments demonstrate that MsLW consistently outperforms existing SOTA watermarking methods under both destructive and regenerative attacks while maintaining superior perceptual quality and generative fidelity. The source code and additional visualization results can be found at https://github.com/chenpian/MsLW.

Generative image steganography has attracted significant attention for its exceptional resistance to steganalysis. However, current generative steganography methods still face limitations in terms of the lack of provable security guarantees under statistical analysis and vulnerability to real-world, unforeseen channel attacks. To address these issues, this paper proposes a novel generative image steganography framework that leverages the Latent Diffusion Model (LDM). Notably, we have uncover a consistent trend: regardless of whether an image has undergone attacks such as compression or noise addition, the sign pattern of values in its latent vector encoded by the LDM remains largely invariant. Capitalizing on this trend, we have devised an adaptive distribution-preserving mapping (ADPM) mechanism, capable of converting a secret message into a latent vector that follows standard normal distribution in an adjustable way. Since both the secret latent vector and the latent vector randomly generated during regular image generation follow the same distribution, satisfying the optimal input conditions for the diffusion model, the proposed method can achieve provable security. Experimental results demonstrate the outstanding performance of our approach in terms of robustness, security, and extraction accuracy.

Visual Large Models (VLMs) are revolutionizing passenger and cargo transportation through autonomous driving (AD), but their immense commercial value also exacerbates the risk of intellectual property (IP) infringement. Model watermarking, a primary IP protection method for deep learning assets, inadvertently degrades model performance, which is unacceptable for safety-critical AD applications. Moreover, traditional IP verification involving third parties introduces vulnerabilities to data and model leakage, potentially jeopardizing both sensitive transportation data and IP of the AD system itself. To tackle aforementioned challenges, this paper introduces a novel IP protection framework for VLMs leveraging zero-knowledge proof technology. Initially, to avoid performance degradation of VLMs in perceiving complex traffic scenarios, this paper proposes a model fingerprinting method incorporating a prior-knowledge-free sample discrimination module. This module quantifies the correlation between samples and decision boundaries using cross-entropy loss, assigning higher weights to high-discriminability samples situated near these boundaries, and fuses the weight matrix with the sample correlation matrix to generate an enhanced fingerprint. This significantly improves the distinguishability between stolen and irrelevant models. Following this, the paper presents an IP verification protocol called zk-DeepIP for VLM, underpinned by zero-knowledge proof, ensuring robust security for AD systems the verification process while remaining compatible with existing IP verification methods. Finally, this paper evaluates the performance of our proposed model fingerprinting technique on CIFAR-10 datasets, where experimental results reveal that the average AUC score improves by 0.03 across various attacks, datasets, and models. Furthermore, in transfer learning scenarios on CIFAR-100 and CIFAR10-c, the average Area Under the Curve (AUC) score increases by 0.12. Notably, the zk-DeepIP protocol effectively safeguards the privacy of both perception models and their associated test cases, ensuring the integrity and security of IP.

Generally, there are two popular ways to protect image copyright, i.e., proactive protection (preventing illegal use via adversarial perturbation) and passive protection (verifying ownership by digital watermarking). However, since the perturbation and watermark embedded into an image will interfere with each other, directly embedding them into the image cannot achieve the proactive protection and passive protection, simultaneously. To address this issue, we propose an image copyright dual-protection approach, which embeds an Extractable and Imperceptible Adversarial Watermark (EIAW) in the image frequency-domain. Specifically, the adversarial watermark is automatically embedded and optimized in the manner of allowing for effectively attacking the Deep Neural Networks (DNNs) and accurately extracting the embedded watermark, simultaneously. Moreover, instead of using the pixel-domain constraints, i.e., Lp norms, we introduce a frequency-domain constraint to optimize the watermark embedding locations. Experiments on ImageNet and CIFAR-10 demonstrate that the proposed EIAW achieves high attack effectiveness (up to 100%) and extraction accuracy (up to 93%), while maintaining good watermark imperceptibility.

The pervasive spread of fake news, particularly through manipulated images, presents a consequential negative impact on society. To prevent fake news images from misleading the public, existing methods focus on verifying the authenticity of news images but ignore source traceability, leaving a gap in creating a complete forensic chain for reliable fake news detection. To simultaneously achieve the goals of authenticity verification and source tracing, we propose a proactive image tagging approach based on a design of Disentangled Invertible Neural Networks (DINN). It can simultaneously embed the dual-tags, i.e., authenticable tag and traceable tag, into each news image prior to publication, allowing for separate extraction for authenticity verification and source tracing. Within the proposed DINN, we design a parallel Feature Aware Projection Module (FAPM) to assist DINN in preserving essential tag information, thereby improving extraction accuracy. In addition, we introduce a Distance Metric-Guided Module (DMGM) that learns asymmetric one-class representations, enabling the dual-tags to exhibit different robustness performances under malicious manipulations. Extensive experiments on diverse datasets and unseen manipulations demonstrate that the proposed tagging approach achieves promising performances on both authenticity verification and source tracing for reliable fake news detection and outperforms the prior works.

Image encryption techniques generally encrypt a secret image into a meaningless noise-like format, which could easily attract attention from attackers who then may try to crack it. On the other hand, image steganography typically embeds secret image data within a cover image, but it inevitably incurs a lot of distortion to the cover image. This makes the secret image data vulnerable to attacks by steganalysis tools. In light of the above, in this paper, we propose a Visually Meaningful Image Encryption (VMIE) scheme via image-to-image reversible transformation based on the Glow model. In this scheme, a secret image is encoded and compressed as a latent vector by the deep compression autoencoder. Then, the latent vector is scrambled and inputted into the Glow model to generate a visually meaningful encrypted image. Extensive experiments demonstrate that the proposed VMIE scheme not only provides desirable security against attacks, but also enables the reconstruction of the original images with negligible quality loss.

Cloud services have been commonly leveraged to store and manage the exponential growth of images, yet this also comes with critical data privacy concerns. Reversible data hiding over encrypted images (RDH-EI) techniques can embed data into encrypted images and support lossless recovery, which can provide an effective solution for securely managing private images in the cloud. However, existing schemes generally suffer from low embedding capacity. Moreover, most of them rely on a single cloud server, which introduces a single point of failure. In this paper, we first propose a pixel correlation recovery (PCR) technique for restoring the pixel correlation excessively disrupted during encryption. Using the PCR technique, we develop a secure $(r, n)$(r,n)-threshold RDH-EI scheme with large embedding capacity and avoidance of single point of failure. In our scheme, a content owner encrypts a confidential image into $n$n shares and distributes them across $n$n independent cloud servers. We design a new encoding method enabling each cloud server to efficiently encode the share, preserving capacity for data embedding. An authorized receiver can later extract the embedded data and reconstruct the confidential image from $r$r shares. Experiments demonstrate that our scheme achieves significantly larger embedding capacity over state-of-the-art schemes.

The rapid growth of AI systems has been fueled by large-scale human data, intensifying concerns over the unauthorized use of intellectual property and privacy-sensitive content during model training. Auditing such misuse is particularly challenging since mainstream AI services operate as black boxes, exposing only generated outputs while concealing their training and inference processes. In this work, inspired by chemical isotope tracing, we introduce the concept of information isotopes to trace training data within opaque AI systems. We propose an information-isotope tracing framework that selectively marks target data elements and detects their propagation in model outputs, providing concrete evidence of data utilization under black-box access. Experiments on thirteen AI models across six datasets demonstrate that our method distinguishes training from non-training data with up to 99% accuracy and strong statistical significance (p < 0.01) using approximately 4,000 words of evidence. An open-source tool is released to support practical data rights protection.

The rapid progression of deep learning technologies has enabled deepfake techniques to generate astonishingly realistic facial images, posing significant threats to unauthorized users and political stability. Currently, defenses against deepfakes primarily focus on detection methods, an ex-post forensic approach that lacks the capability to prevent malicious activities. Additionally, the majority of proactive defense strategies attempt to disrupt deepfake models by leveraging adversarial perturbations. However, these perturbations have proven susceptible to lossy compression, compromising their robustness across diverse models. Once these perturbations fail, proactive defense systems likewise collapse. To bridge this gap, this paper proposes a robust dual-pronged proactive defense framework against Deepfakes, named Adversarial Semi-Fragile Watermarking (ASFW). Primarily, recognizing the constraints of purely passive detection or proactive defense strategies, this paper presents a dual-branch proactive defense framework. One branch utilizes a generator to create adversarial semi-fragile watermarking that disrupt deepfake models for proactive defense. The other branch integrates a detector to authenticate the integrity of these watermarks, enabling the identification of image manipulation. Therefore, even if proactive defense fails, our method retains the capability of conducting passive forensics through verifying the integrity of adversarial semi-fragile watermarks, offering a more comprehensive shield against deepfakes. Moreover, in order to enhance the robustness of ASFW against lossy compression, we incorporate a differentiable compression module during training. Extensive experiments demonstrate that our proposed method excels in terms of visual quality, cross-model capability, and robustness, proving its effectiveness.

Abstract Integrating multi-source visual data is essential for environmental perception but poses significant challenges for cognitive-level understanding due to cross-modal ambiguity. This study introduces the Deep Fuzzy K-Means with Multi-source Consistency (DFKMC) framework, which com-bines deep representation learning with fuzzy clustering under dual consistency constraints. A local consistency principle selects high-confidence samples, while a global consistency loss aligns the clustering process with semantic distributions. Extensive experiments on six benchmarks show DFKMC’s superior performance over state-of-the-art methods in clustering accuracy, normalized mutual information, and purity. The proposed consistency modules are proven indispensable via ablation studies. DFKMC provides a robust, end-to-end solution for deriving interpretable cognitive assignments from heterogeneous visual data.

This letter proposes a novel scheme of single-antenna channel modulation with variable channel-path (SACM-VCP) for single-input multiple-output (SIMO) systems. Using a switch controller, the SACM-VCP dynamically activates either one or two channel paths via radio frequency (RF) mirrors. This introduces a new switching index dimension that enhances the spectral efficiency without requiring additional RF chains. To further improve the bit error rate (BER) performance, hybrid signal constellations are co-designed with the channel path(s) selection under the constraint of squared minimum Euclidean distance. Moreover, the detection complexity and average bit error probability are analyzed by using the union bounding and moment generating function techniques. Simulation results demonstrate that the SACM-VCP achieves significant SNR gains over existing MBM-based schemes, presenting an improved trade-off between spectral efficiency, reliability, and complexity.

Index modulation, which develops the additional information using the antenna index (AI) domain, is a promising modulation technique for next wireless communications. In view of a rich radio scattering environment around transmit antennas (TAs), utilizing the indexes of channel fade realizations, media-based modulation (MBM), which develops the additional information using the channel index (CI) domain, is a recently potential channel modulation and viewed as a potential key technique for enhancing the reliability of communication systems. In this paper, to simultaneously carry the additional information including the AI and CI information bits, the integration of index modulation and MBM is investigated. Firstly of all, the application of the MBM technique to the extended space index modulation (ESIM) system, called as ESIM-MBM, is investigated to improve the spectral efficiency (SE) and error performance of the multiple-input multiple-output with index modulation (MIMO-IM) systems. Specifically, after obtaining one ESIM vector by modulating the real and imaginary parts of one mapped symbol on one or two active TAs using the selected AI vector, according to the number of active TAs, the real and imaginary parts of one mapped symbol are respectively performed Kronecker products by two selected CI vectors with two subparts of CI bits, result in the transmitted MBM vector. Secondly, in order to further extend the size of signal spaces to improve the AI information, with the combination of one or two signal points from QAM and secondary QAM constellations, a new design of enhanced signal spaces with two active TAs assisted the MBM system (ESS-TTAs-MBM) is proposed to enhance the throughput of the communication system. Furthermore, using the maximum likelihood (ML) at the receiver, the comparisons of detection complexities of the proposed ESIM-MBM and ESS-TTAs-MBM with different MBM-based IM schemes are analyzed, and the theoretical average bit error probability (BEP) is also formulated and shown to match well with the Monte-Carlo simulation results at different TAs and SEs in the high SNR region. Finally, a significant improvement in the SE and bit error performance of the proposed schemes is demonstrated with other classic MBM-based IM schemes such as quadrature spatial modulation assisted MBM (QSM-MBM) and quadrature channel modulation (QCM).

Steganography is one of most important techniques for covert communication. In recent years, generative steganography, which transforms a secret information into a generated picture, is a prospective steganography-resistant technique. Nevertheless, it is difficult to achieve a good trade-off between information hiding ability and extraction accuracy because of the low efficiency and irreversibility of the secret-to-picture conversion. In order to solve this problem, this paper proposes a secret message-driven picture generation solution for generative steganography. The presented SM-IG scheme is founded on the design of a stable version of the Nearly Independent Component Estimation (Stable NICE) model, allowing for a stable bijection mapping between a potential space with simple distributions and an picture space with complex distributions. During the secret to picture conversion, a latent vector is constructed, driven by a given secret message, which is then mapped to the generated picture via the Stable NICE modelAs a result, the secret information is eventually converted into the generated picture. Due to the good efficiency and reversibility of the SM-IG scheme, this steganography method has high hiding capability and accurate message extraction accuracy. The experiments prove that the proposed SM-IG can simultaneously realise good-level hiding capacity (as much as 4 bpp) and precise extraction accuracy (close to 100 % accuracy) without compromising the required resistance to detection and imperceptibility.

Steganalysis aims to detect the presence of concealed information within seemingly normal carriers in network transmissions, playing a crucial role in maintaining cybersecurity. With the rapid development of social networks, steganalysis techniques targeting social network texts have attracted significant interest from researchers in recent years. However, existing steganalysis techniques for social texts generally focus on analyzing the statistical features of the text itself, neglecting the relational features between texts, thereby limiting their detection capabilities. In this paper, we propose a novel text steganalysis feature enhancement method—TGCA. This method considers the relational features between texts by introducing GNN, while utilizing Transformers to expand the receptive field of GNNs and incorporating a cross-attention mechanism to reduce the aggregation of noise, thus mitigating the inherent limitations of GNNs. As a result, TGCA more effectively extracts and integrates textual and topological features, enhancing the model's performance in detecting steganographic texts. Experimental results demonstrate that TGCA outperforms existing methods by better leveraging graph and textual features to identify malicious steganographic texts. Our code is available at https://github.com/PandaaKai/TGCA.

Existing Automatic Speaker Verification (ASV) systems show inferior performance in dealing with the cross-domain challenges posed by far-field utterances. To mitigate domain mismatch issues in far-field scenarios, this paper proposes a novel but promising Multi-layer Knowledge Distillation-based Far-field Speaker Verification (MKD-FSV) system. MKD-FSV employs a teacher-student framework with a well-thought-out Multi-layer Knowledge Distillation Strategy (MKDS) to transfer domain-invariant knowledge from close-talking utterances, guiding the student model to effectively learn far-field data. MKDS elaborately incorporates Feature Knowledge Distillation (FKD) with Decoupled Knowledge Distillation (DKD) in a complementary manner to ensure a more comprehensive knowledge transfer and significantly enhance the system's performance. Specifically, FKD captures domain-invariant speaker characteristics more effectively, transferring crucial features from the teacher to the student model to develop a more robust, generalized speaker embedding space, while DKD enhances speaker feature discriminability by balancing information between target and non-target classes, reducing interference, and improving accuracy in cross-domain mismatch scenarios. This dual approach ensures that the student model inherits robust speaker embeddings while refining its decision-making, leading to improved verification performance. Moreover, a reparameterization technique is utilized to reduce model complexity and enhance inference efficiency. Extensive experimental results demonstrate that MKD-FSV outperforms existing methods in far-field ASV tasks, achieving both higher verification accuracy and significantly improved inference efficiency through the reparameterization technique, making it highly applicable in complex real-world scenarios.

With the rapid development of deep learning in the field of computer vision, the performance of core vision tasks such as image recognition has achieved significant improvement. In nighttime environment, due to the low-light condition and reduced visibility, cross-domain transformation of nighttime images based on Generative Adversarial Network (GAN) model can effectively improve the accuracy of nighttime recognition models. However, the existing GAN models are difficult to be effectively deployed on resource-constrained devices due to the requirement of high storage space and computational resource. To this end, this paper proposes a shared attention network based on the attention mechanism with the CycleGAN structure, and designs an online knowledge distillation method to compress and optimize the model, so as to obtain a lightweight model to achieve the nighttime-daytime cross-domain image transformation. Experimental results demonstrate that the proposed model achieves the state-of-the-art performance in the task of Nighttime-Daytime Image Transformation. This is of great significance for edge devices to solve the problem of recognition at night.

In recent years, research on information-hiding techniques based on network streaming media has focused on how to covertly embed secret information within real-time transmissions to achieve clandestine communication. The misuse of such technologies poses significant security risks, such as the dissemination of malicious codes, commands, viruses, and more. The existing methods for steganalysis of network voice streams generally face challenges in universality, exhibiting poor adaptability to steganographic detection scenarios with non-identity distributions. To address these issues, we introduce a framework named the Domain-Aware Enhanced Framework for VoIP Steganalysis (DAEF-VS), which harnesses the CutMix technology to enhance the shared steganographic domain features and employs the Domain-Aware Learning Model to fine-tune these features, thereby significantly improving generalization capabilities. Extensive experimental results demonstrate that our approach vastly surpasses existing advanced methods in terms of universality across a variety of steganographic detection scenarios.

Constructing a deepfake fingerprint detection model (D2FDM) to counteract spoofing attacks stemming from forged fingerprints is a formidable challenge, requiring massive fingerprint samples, computing resources, and profound professional expertise. Consequently, it is of paramount importance to formulate an effective strategy to protect the intellectual property (IP) of D2FDM. To achieve this, this paper proposes a practical ProIP-D2FDM framework to authenticate the D2FDM. Firstly, the identity of model ownership is hidden in the cross-material fingerprint using generative steganography. The generated invisible additive noise serves as a watermark, rendering the backdoor inconspicuous, and facilitating its embedding through collaborative training with the original fingerprint dataset. Then, the key trigger set is constructed by sampling the original fake fingerprints, and the target label of the trigger set is employed as the foundation for IP verification. Once authorized, the embedded identity can also be distributed. Finally, to prevent the functionality of D2FDM from being stolen via distillation attacks, this paper further designs a novel non-distillable strategy. In this strategy, the student model is deprived of any prior knowledge acquisition. Experimental results demonstrate that our method achieves a remarkable 100% watermark verification accuracy and 100% success rate in user identity authentication on D2FDM trained using two public fingerprint datasets, namely LivDet2017 and LivDet2019. When the non-distillable strategy devised in this paper is used, the performance of the student model, with the D2FDM as the teacher model, exhibits a 4% decline. Moreover, the framework presented in this paper shows remarkable robustness against several common attacks.