INTRODUCTION: Online public opinion networks face critical challenges in privacy protection, traceability, and dynamic governance due to their distributed, multi-node propagation nature. Existing approaches either sacrifice data integrity through irreversible encryption or lack the collaborative intelligence needed for large-scale opinion monitoring.OBJECTIVES: This paper proposes STORM, a framework that integrates Multi-Agent Systems (MAS) with Reversible Data Hiding (RDH) to enable privacy-preserving, traceable, and lossless opinion data governance.METHODS: We first conduct a systematic literature synthesis identifying three technology clusters and four research gaps in the MAS-RDH intersection. We then design a five-layer cooperative architecture (Perception, Processing, Hiding, Control, Recovery) with a formal mathematical model. We further detail the multi-agent communication protocol, scheduling mechanism, and collaborative decision-making process. Extensive numerical experiments evaluate STORM across clean-channel capacity (8– 128 bits), Gaussian noise robustness (σ up to 2.0), multiple attack models (cropping, quantization, adversarial perturbation), opinion dynamics scenarios, computational complexity, and statistical indistinguishability (KS test).RESULTS: STORM achieves perfect reversibility (BER=0) across all clean-channel configurations, maintains zero bit error rate under Gaussian noise with standard deviation up to 2.0, and provides fine-grained distortion control via a forgetting coefficient (β). Extended experiments confirm BER=0 up to 96-bit capacity, robustness against adversarial perturbations (BER< 0.0003), and successful embedding in simulated public opinion dynamics (BER=0 across three opinion distribution types). Computational complexity scales as O(N3 + N2K) for precomputation and O(NK) per embedding instance. Compared to existing methods, STORM is the first scheme simultaneously satisfying full reversibility, multi-agent coordination, and high noise robustness. CONCLUSION: The STORM framework offers a new technical paradigm for intelligent, precise, and compliant online opinion security governance, with validated theoretical advantages and clear directions for future deployment on real-world platforms.

Image steganography based on deep neural networks (DNNs) aims to embed secret information into images more rationally. We observe that due to the inherent generalization capability of deep neural network, a third party can extract secret information in certain schemes merely by knowing the model structure. We term this vulnerability the generalization attack-based steganalysis. Furthermore, to address the covert transmission of secret models, we propose a novel steganographic network. To address the aforementioned risks and the covert transmission of secret models, we propose a novel image steganography method with strong constraints against generalization attack-based steganalysis. To resist steganalysis, the model is trained to map natural images to secret images without modifying the original natural images, and a key is used for embedding and extracting the secret model. For deployment, the secret model is embedded into a stego model that performs ordinary tasks, and can later be extracted using a key. Experimental results show that the proposed scheme can effectively resist generalization-attack-based steganalysis. Before and after camouflage, the average maximum differences in task performance degradation of the stego model are 1.16 dB and 0.38 dB in the Eval and No Eval modes, respectively, which can meet the requirements of covert transmission for secret tasks. For the reconstructed secret images, the proposed scheme achieves good performance in terms of evaluation metrics such as PSNR, MAE, and RMSE.

The rapid progress of generative artificial intelligence (GenAI), especially generative adversarial networks (GANs), has stimulated growing interest in learning-based pseudorandom number generation. However, most existing GAN-based PRNGs are evaluated primarily through empirical statistical batteries, providing limited formal assurance against computationally powerful adversaries and lacking reductions to established hardness assumptions. This paper proposes a PRNG construction that integrates a GAN-based Bernoulli noise expansion sampler into an LPN-based framework. We formalize a neural-structured LPN variant (GAN-LPN) that captures practical non-idealities of learned sampling via explicit, measurable deviation parameters, and we carry these terms through the security analysis as additive degradation in the final advantage bound. We then present a game-based reduction showing that, under the stated assumptions and deviation bounds, distinguishing the PRNG output from random can be reduced to the hardness of the corresponding underlying LPN-type problem. Experimental results show that the generated sequences pass NIST SP 800-22 and additional empirical benchmarks, including the Hamming Distance test and a practical next-bit prediction test, providing supplementary evidence on the statistical quality of the output.

Reversible data hiding in encrypted images (RDHEI) enables secret data to be embedded into encrypted images while allowing exact recovery of the original image. However, existing reserving room before encryption (RRBE)-based methods do not fully exploit bit-plane redundancy and generally provide limited support for fine-grained recovery authorization. To address these limitations, this paper proposes a high-capacity RRBE-based RDHEI scheme that integrates dual-stage bit-plane compression with block rearrangement. In addition, a ciphertext-policy attribute-based encryption (CP-ABE)-based key encapsulation mechanism is introduced to support fine-grained and policy-driven image recovery. First, pixel predictors are employed to generate compact prediction errors. Then, consecutive all-zero high-order bit planes are encoded to exploit global sparsity, while the remaining bit planes are further compressed using adaptive hierarchical block variable-length coding (AHBVLC) to capture local structural redundancy. Finally, the compressible bit planes are reorganized, and the image blocks are rearranged according to their embedding capacities. Experiments on BOSSBase, BOWS-2, and UCID demonstrate that the proposed method achieves competitive embedding performance while enabling exact recovery of the original image. The proposed scheme also supports flexible, policy-driven image recovery in multi-user scenarios.

reversible data hiding methods in encrypted images (RDH-EI) are primarily designed for point-to-point scenarios and not suitable for multiparty communication. To address this issue, a novel RDH-EI scheme based on multi-key (k, n)-threshold decryption (RDH-EITD) is proposed, in which a dual-phase embedding method is designed to support authentication for both the content owner and central server. In RDH-EITD, the private key of Paillier is split into n shares, which are then allocated to n distributed receivers. The image is first encrypted and then embedded with additional data to generate the marked ciphertext, which is uploaded to data server. The server can perform the 2nd-phase embedding. The dual-marked ciphertext is then distributed to n receivers. Each receiver can extract the 2nd-phase embedded data but cannot decrypt the image individually. Only when k out of n receivers submit their partially decrypted results, can the image be decrypted. Then the 1st-phase embedded data can be extracted and the original image can be recovered losslessly. Security reduction is employed to formally prove the semantic security of RDH-EITD. Experimental results demonstrate that RDH-EITD preserves (k, n)-threshold decryption, enabling resistance against k-1 collusion attacks and tolerance of n-k failures. The dual-phase embedding achieves embedding rates of lambda - 9 bpp (bits per pixel) and 14 bpp respectively with security parameter lambda. For an image of length L in one-to-n communication scenarios, the time complexity of RDH-EITD reaches O(L(lambda(3) + k(2))), outperforming existing Paillier-based point-to-point solutions.

Existing reversible data hiding in encrypted images (RDH-EI) methods face limitations in embedding layers, capacity, and quantum resistance, making them unsuitable for hierarchical multiparty scenarios. This paper proposes a novel multi-layer RDH-EI scheme based on Ring Learning with Errors (RLWE), leveraging RLWE's additive homomorphism and designing a polynomial coefficient partitioning strategy. The content owner first encrypts the original image using RLWE, and multiple data hiders sequentially embed secret messages into the ciphertext through homomorphic addition. The receiver can decrypt the marked ciphertext to extract all embedded messages and recover the original image losslessly. Experimental results demonstrate that the proposed method supports up to n-1 embedding layers with security parameter n and achieves an embedding rate (ER) of 8 to 2040 bits per pixel (bpp). Compared to state-of-the-art RDH-EI approaches, this scheme offers significant advantages in embedding layers, capacity, and quantum resilience, providing a practical solution for secure hierarchical multiparty communication.

Mobile and IoT visual devices increasingly process facial images under resource constraints, creating privacy risks during storage, transmission, and cloud processing. Existing reversible face anonymization methods primarily emphasize correct-key recovery, with limited attention to mobile deployment and wrong-key security. This paper proposes Mobile-SST-RIR, a resource-aware face privacy framework for IoT visual devices with key-conditioned reversible recovery. Following a source-side protection paradigm, raw facial images are transformed into privacy-preserving representations before transmission or storage, while identity recovery is permitted only with a valid key. To accommodate resource-constrained devices, privacy generation is performed through a lightweight feed-forward process at the source, whereas key validation and image reconstruction are delegated to trusted edge or cloud environments. Mobile-SST-RIR combines multi-scale AdaIN residual blocks, gated skip connections, and a FiLM-after-AdaIN key injection strategy to balance identity obfuscation, visual utility, and authorized recoverability. A Key Validator and wrong-key constraint are further introduced to strengthen access control and suppress unauthorized reconstruction. Experiments on CelebA show that the proposed method achieves high-quality correct-key recovery, with a PSNR of 29.47 dB, SSIM of 0.869, and LPIPS of 0.040. The identity similarity between protected and original images is reduced to 0.295, while wrong-key reconstruction remains low at 0.344. The Key Validator attains a matching accuracy of 0.964, and security evaluations confirm robustness against black-box and limited-query attacks. These results indicate that Mobile-SST-RIR provides a practical balance among source-side biometric privacy protection, authorized recovery, visual usability, and wrong-key risk control for mobile and IoT visual applications.

Deep steganography utilizes the powerful capabilities of deep neural networks to embed and extract messages, but its reliance on an additional message extractor limits its practical use due to the added suspicion it can raise from steganalyzers. To address this problem, we propose StegaINR, which utilizes Implicit Neural Representation (INR) to implement steganography. StegaINR embeds a secret function into a stego function, which serves as both the message extractor and the stego medium for secure transmission on a public channel. Recipients only need to use a shared key to recover the secret function from the stego function, allowing them to obtain the secret message. Our approach employs continuous functions, enabling it to handle various types of messages. To our knowledge, this is the first work to introduce INR into steganography. We perform evaluations on image, climate data, and Neural Radiance Field (NeRF) synthetic dataset to test our method in different deployment contexts.

Implicit neural representations (INRs) enable continuous, resolution-independent image modeling and have recently been explored for multi-image steganography. However, many INR-based schemes hide secrets by modifying a small subset of network parameters, making embedded information fragile when the model is quantized or pruned. To improve robustness, we propose a dual-key INR framework that embeds secret information into the input encoding stage rather than directly modifying network weights. Specifically for each secret image, we generate a pair of keys that induce frequency perturbation and phase shift in Fourier feature encoding, so multiple images can be represented by a single shared INR while remaining independently recoverable. We further integrate simulated quantization and pruning into training in a progressive manner to enhance tolerance to parameter perturbations in practical deployment. Experiments demonstrate that the proposed method maintains stable reconstruction quality under low-bit quantization as well as structured and unstructured pruning, while supporting multi-image embedding with competitive reconstruction fidelity.

To address the limited channel robustness of invertible neural networks against Gaussian noise, JPEG compression, and scaling/resampling in image information hiding, while maintaining their embedding capacity, we propose an invertible steganography algorithm that integrates a dual-branch attention mechanism in the frequency domain. First, a dual-branch attention module is designed to adaptively generate embedding weights from both global subband energy and local texture pixel dimensions after wavelet multi-frequency decomposition, thereby constructing a four-channel extended secret payload. Second, a four-channel extension strategy is developed to enhance embedding capacity without introducing random noise distortion. Additionally, dual enhancement mechanisms are employed to preprocess the original carrier before embedding, optimize frequency-domain feature distribution, and restore images contaminated by noise, compression, or scaling attacks during reconstruction. Experimental results demonstrate that under a 1:1.34 embedding ratio, the proposed algorithm achieves a clean-channel C-PSNR of 36.27 dB; under Gaussian noise (σ=10) and JPEG(QF=50) attacks, PSNR values for all channels reach 27.96 dB and 29.43 dB respectively. Compared to comparable models without attention mechanisms, this approach improves carrier quality and secret recovery metrics by over 2 dB on average, combining high transmission capacity with excellent channel robustness and visual imperceptibility.

In multi-agent systems (MAS), semantic strategies undergo decomposition, transformation, and composition across distributed agents, exposing them to eavesdropping, tampering, and provenance loss. Conventional encryption disrupts decisiondata formats, whereas standard information hiding cannot losslessly restore the carrier — both undermine high-precision coordination. This paper proposes a hierarchical Reversible Data Hiding (RDH) framework for trustworthy semantic-strategy coordination. We build a semantic — carrier feature alignment model grounded in a formally defined alignment score, and design three-layer RDH primitives — text histogram-shifting at the semantic layer, topology-aware embedding at the coordination layer, and lightweight prediction-error expansion (PEE) at the execution layer — each with explicit carrier forms and distortion constraints. A formal threat model with concrete adversary capabilities bounds the security guarantees. A capacity distortion — accuracy multi-objective optimization and distributed Byzantine-fault-tolerant cross-validation enable autonomous self-healing. Experiments on a simulated UAV swarm (3 – 11 agents, grid topology, DDPG policy, 500-bit payload) show 99.8% extraction accuracy, <3% decision degradation, and sub-20 ms round-trip latency, outperforming PEE-RDH, ciphertext-domain RDH, covert communication, Secure-MARL, and blockchain-watermarking baselines.

With the widespread application of neural radiance fields (NeRF) in three-dimensional scene representation and rendering, protecting these models from unauthorized use has become critical. We propose a NeRF watermarking algorithm based on model redundancy, using a pre-training watermark embedding strategy to embed a watermark NeRF model into the carrier NeRF model for copyright protection. First, on an initial implicit representation network of the watermark NeRF, a random neuron inactivation is applied, and the implicit network is trained using a set of image watermark data. Second, neuron pruning techniques are used to remove unimportant weights in the watermark NeRF network, with pruning positions recorded as keys. Finally, a NeRF model for carrier representation is trained while fixing important weights, achieving NeRF watermark embedding. Copyright verifiers can extract the watermark network from the carrier NeRF network using keys to prove ownership. Experimental results show that our method not only maintains high-quality reconstruction effects and achieves lossless extraction of watermarks but also that the watermark type with implicit representation has high robustness and can resist various common attack methods.

To address the limited embedding capacity of existing reversible data hiding in encrypted images (RDH-EI) schemes, this paper proposes a high-capacity dual-layer RDHEI algorithm based on ElGamal encryption. By exploiting the redundancy of the plaintext space and homomorphic property of ElGamal, the proposed scheme enables dual-layer embedding: the content owner preprocesses the image and embeds initial data before encryption, while the data hider embeds additional messages in the ciphertext domain. Receivers can extract both layers of embedded data and fully recover the original image. Experimental results demonstrate embedding capacities of 1016 bpp (bits per pixel) and 10 bpp for the two layers. And security reduction provides formal semantic security proof.

Traditional steganography conceals information by modifying cover data, but steganalysis tools easily detect such alterations. While deep learning-based steganography often involves high training costs and complex deployment. Diffusion model-based methods face security vulnerabilities, particularly due to potential information leakage during generation. We propose a fixed neural network image steganography framework based on secure diffusion models to address these challenges. Unlike conventional approaches, our method minimizes cover modifications through neural network optimization, achieving superior steganographic performance in human visual perception and computer vision analyses. The cover images are generated in an anime style using state-of-the-art diffusion models, ensuring the transmitted images appear more natural. This study introduces fixed neural network technology that allows senders to transmit only minimal critical information alongside stego-images. Recipients can accurately reconstruct secret images using this compact data, significantly reducing transmission overhead compared to conventional deep steganography. Furthermore, our framework innovatively integrates ElGamal, a cryptographic algorithm, to protect critical information during transmission, enhancing overall system security and ensuring end-to-end information protection. This dual optimization of payload reduction and cryptographic reinforcement establishes a new paradigm for secure and efficient image steganography.

Recently proposed steganalysis methods based on the local optimality of motion vector prediction (MVP) indicate that the existing HEVC (high efficiency video coding) motion vector (MV) domain video steganography algorithms can disturb the optimality of MVP in advanced motion vector prediction (AMVP) technology. In order to improve the security of steganography algorithm, this paper proposes an MV domain steganography method in HEVC based on MVP's index and motion vector difference (MVD). First, we analyze the conditions that need to be met for steganography to resist attacks from MVP's optimality features and other traditional steganalysis features. Then, a distortion function for minimizing embedding distortion is designed, and an algorithm for secret message embedding and extraction in units of inter-frame is proposed. Experimental results show that the proposed algorithm can resist attacks based on the optimality of MVP and also has high security against other traditional steganalysis methods. In addition, the proposed algorithm has excellent performance in visual quality and coding efficiency, and can be applied to practical scenarios of video covert communication.

Symmetric encryption algorithms learned by the previous proposed end-to-end adversarial network encryption communication systems are deterministic. With the same key and same plaintext, the deterministic algorithm will lead to the same ciphertext. This means that the key in the deterministic encryption algorithm can only be used once, thus the encryption is not practical. To solve this problem, a nondeterministic symmetric encryption end-to-end communication system based on generative adversarial networks is proposed. We design a nonce-based adversarial neural network model, where a “nonce” standing for “number used only once” is passed to communication participants, and does not need to be secret. Moreover, we optimize the network structure through adding Batch Normalization (BN) to the CNNs (Convolutional Neural Networks), selecting the appropriate activation functions, and setting appropriate CNNs parameters. Results of experiments and analysis show that our system can achieve non-deterministic symmetric encryption, where Alice encrypting the same plaintext with the key twice will generate different ciphertexts, and Bob can decrypt all these different ciphertexts of the same plaintext to the correct plaintext. And our proposed system has fast convergence and the correct rate of decryption when the plaintext length is 256 or even longer.

To address the issues of image quality degradation and file size expansion encountered during reversible data hiding (RDH) of JPEG images, a JPEG reversible data hiding algorithm based on block smoothness estimation and optimal zero coefficient selection is proposed. Firstly, a block smoothness estimation strategy is designed based on the number of zero coefficients and non-zero quantisation table values within DCT blocks, prioritising DCT blocks with higher smoothness for information embedding. Subsequently, under a given embedding payload, an optimal zero coefficient selection strategy is introduced. Blocks are partitioned into embedding regions and non-embedding regions based on a preset position threshold T. Within embedding regions, the frequency of zero coefficients at different positions across all blocks is statistically analysed, with embedding prioritised at positions exhibiting the highest zero coefficient frequency to enhance embedding efficiency. Concurrently, by setting positive and negative displacement gaps to constrain the modification range of non-zero coefficients, invalid shifts are minimised. This further enhances visual quality while controlling file expansion. Experimental results demonstrate that, compared to existing algorithms, the proposed method achieves a peak signal-to-noise ratio improvement of 0.75 to 3.62 dB under fixed embedding capacity. File expansion is reduced by 1038 to 2243 bits, whilst enabling fully reversible image restoration.

Until now,some reversible data hiding in encrypted images(RDH-EI)schemes based on secret sharing(SIS-RDHEI)still have the problems of not realizing diffusivity and high embedding capacity.Therefore,this paper innovatively proposes a high capacity RDH-EI scheme that combines adaptive most significant bit(MSB)prediction with secret sharing technology.Firstly,adaptive MSB prediction is performed on the original image and cryptographic feedback secret sharing strategy encrypts the spliced pixels to spare embedding space.In the data hiding phase,each encrypted image is sent to a data hider to embed the secret information independently.When r copies of the image carrying the secret text are collected,the original image can be recovered lossless and the secret information can be extracted.Performance evaluation shows that the proposed method in this paper has the diffusivity,reversibility,and separability.The last but the most important,it has higher embedding capacity.For 512 ×512 grayscale images,the average embedding rate reaches 4.7358 bits per pixel(bpp).Compared to the average embedding rate that can be achieved by the Wang et al.'s SIS-RDHEI scheme,the proposed scheme with(2,2),(2,3),(2,4),(3,4),and(3,5)-threshold can increase by 0.7358 bpp,2.0658 bpp,2.7358 bpp,0.7358 bpp,and 1.5358 bpp,respectively.

Federated learning (FL) allows for collaborative training without centralizing data, but concerns regarding model privacy leakage, intellectual property theft and poisoning attacks have hindered its development. To mitigate such risks, this article proposes embedded message authentication code technology (EMAC) to integrate encryption, digital signatures, and watermark functions for model security. In EMAC, the authentication data is embedded into the model ciphertext using reversible data hiding after encryption. The marked ciphertext supports data extraction for subsequent authentication and lossless decryption for testing and training simultaneously. Based on EMAC, a novel FL with security authentication and traceability of poisoning (FL-SATP) is proposed, which integrates privacy protection, identity authentication and poisoning traceability into FL. The poisoner tracing is designed to detect and identify poisoners retrospectively based on the practical performance of trained or aggregated models, thus removing the malicious users' model and deterring poisoning behaviors. Theoretical analysis and experimental results demonstrate that FL-SATP could ensure the confidentiality of the model content, the availability of model function, and that when more than half of the users are benign, the proposed method can accurately and efficiently pinpoint all malicious poisoners in FL.