Graph data publication is of significant value in areas such as social network analysis and recommendation systems. However, under low privacy budgets, existing differential privacy (DP) graph generation methods tend to introduce excessive noise, leading to structural distortion and degraded utility. To address this, we propose PrivDPR-CD, a novel differential privacy graph publishing framework that employs a Deep PageRank (DPR) learning mechanism to robustly capture and preserve community structures. Specifically, our method achieves more stable DPR learning through layer-wise differential privacy budget allocation and leverages privacy-preserving graph structures for community partitioning to mitigate reconstruction noise. Subsequently, an improved Expected Configuration Model is utilized to generate a high-fidelity graph. Experimental results on three real-world datasets demonstrate that PrivDPR-CD significantly enhances community preservation, node importance maintenance, and topological consistency, outperforming state-of-the-art methods, particularly under low 𝜖 (i.e., stringent privacy) conditions.

The spatial crowdsourcing (SC) can quickly assign tasks and obtain feedback based on task requirements and workers' locations, which brings great convenience to task assignment. However, sensitive information can also be easily obtained by SC platforms. To prevent information leakage, various privacy-preserving task assignment schemes have been proposed. However, existing schemes have low query efficiency and may leak pattern privacy, task content, or worker preference. To address the above challenges, this article proposes a differentiated privacy-preserving task assignment scheme based on generative adversarial networks in spatial crowdsourcing (DPGAN-SC). This scheme leverages generative adversarial networks (GANs) to generate disguised locations for both tasks and workers, which are then used in the task-matching process. Within the standard area range, no location can be distinguished, ensuring location privacy while preventing adversaries from analyzing search patterns through matching results. The combination of location disguise and task content encryption makes it impossible for adversaries to infer worker preferences and access patterns through the matching process. In addition, to meet differentiated privacy requirements, DPGAN-SC leverages GANs to design a three-level privacy-classification mechanism. This mechanism categorizes private data while minimizing unnecessary privacy overhead. Compared with existing schemes, DPGAN-SC improves query efficiency by 100 times while ensuring the comprehensive privacy preservation.

With the rapid development of edge computing and the explosive growth of image data generated by IoT and mobile devices, an increasing number of users prefer to perform privacy-preserving image storage and retrieval tasks directly at the edge. However, existing solutions typically rely on basic encryption methods and shallow feature extraction, leading to inadequate data security and poor retrieval performance. In this paper, we propose a Dynamic Multi-Stage Encryption (DMSE) method combined with a semantically rich fusion feature to achieve high-precision and privacy-preserving image retrieval in edge environments. Specifically, the proposed method first divides the image into blocks and applies random shuffling, followed by channel and pixel-level XOR encryption to generate a hybrid encrypted image. Then, we extract global features from the encrypted image using the histogram of Discrete Cosine Transform (DCT) coefficients. In addition, a multi-scale convolution block is designed to extract stable and robust local features under encryption. Finally, deep learning is utilized to fuse the global and local features, capturing both the holistic structure and fine-grained semantics of the image. This comprehensive feature representation significantly improves retrieval accuracy while ensuring privacy. Extensive experiments validate that our approach outperforms existing methods in both security and retrieval effectiveness, making it well-suited for edge computing scenarios with limited resources and high privacy demands.

With the rapid development of cloud computing security, multi-user conjunctive dynamic searchable encryption is receiving more and more attention. However, existing schemes mainly rely on the XSet structure to implement conjunctive queries, which brings significant additional computational costs such as exponentiation, an extra round of communication. In addition, the leakage of keyword pair result pattern (KPRP) has been overlooked. In this work, a practical multi-user conjunction dynamic searchable encryption scheme (named MORF-MUDSE) is proposed, which can against a passive attack by a server colluding with any search user while ensuring the two basic attributes of both forward and backward security. Specifically, a mask-obfuscation random function protocol (named MORF) that includes document keyword set information is designed to implement a new subset validation method of conjunctive queries. This method only completes one round of communication to obtain results, which greatly improves computational efficiency while protecting KPRP. Moreover, a delayed update strategy is proposed, which address the issue of information synchronization between data owner and users during those updates. Through theoretical and experimental analysis, it has been shown that the proposed scheme in this paper has better efficiency in computation and communication compared to the state-of-the-art schemes.

With the widespread adoption of graph-structured data, protecting the complex relational information between nodes and edges while preventing sensitive information leakage has become a critical challenge. However, existing edge protection methods either introduce noise directly into the adjacency matrix, resulting in significant information loss, or uniformly apply noise across all edges, leading to imbalanced privacy budget allocation and inefficiency. To address these issues, we propose DPEI, a Differential Privacy-based Edge Information protection solution designed to safeguard the edge relationships between two nodes, thus reducing the risk of privacy leakage and preventing attackers from repeatedly inferring internal community relationships from the released graph data. Specifically, DPEI achieves protection through PPO (Proximal Policy Optimization)based selection of locally optimal thresholds combined with adaptive Laplace noise operations, and attachment nodes below the threshold into high-information edges to enhance relational information protection. Subsequently, unlike traditional uniform allocation, DPEI distributes the privacy budget in proportion to the information content of each edge, ensuring that edges with higher information content receive stronger privacy protection. Extensive experiments conducted on three real-world graph datasets demonstrate that DPEI significantly outperforms existing methods across seven commonly used graph metrics, thereby validating its effectiveness and practicality.

The Underwater Internet of Things (UIoT) faces critical challenges in topology construction due to harsh acoustic environments and limited node energy, typically resulting in a trade-off stalemate between energy balancing and topological robustness. To address this, a topology control algorithm based on Degree-Constrained and Multi-dimensional Link Utility assessment (DC-MLU) is proposed. Specifically, a multi-dimensional utility model is established to quantitatively evaluate link quality by integrating residual energy, communication energy overhead, propagation delay, and transmission reliability. Furthermore, a spatial density-driven dynamic degree constraint mechanism is designed, enabling sensor nodes to adaptively adjust their connectivity according to the local neighborhood scale. Simulation results demonstrate that, compared to algorithms such as EDNPM, EEUMC, and CTSO, the proposed algorithm effectively suppresses redundant links, significantly extends the overall network lifetime, and achieves dual-performance in both structural robustness and energy balancing.

The Internet of Vehicles (IoV) is a crucial component of Intelligent Transportation Systems, enhancing the intelligence and connectivity of transportation networks through comprehensive data support and communication capabilities. However, since each communication entity within the IoV operates in an open network environment, it is vulnerable to various threats, including user data leakage and vehicle operational security issues. Moreover, the IoV is highly time-sensitive, necessitating the design of a secure and efficient authenticated key agreement (AKA) protocol in IoV. Unfortunately, most current solutions only achieve one-to-one authentication and lack scalability for one-to-many or cross-domain scenarios. As a result, their efficiency remains limited, while alternative AKA schemes that support these use cases tend to introduce linearly increasing costs and computational overhead. To address these challenges simultaneously, we propose BCEDAS that utilizes chameleon hash function, physical unclonable functions and blockchain technology, which minimizes reliance on centralized trusted party while meeting essential security features, including unlinkability, non-repudiation, and side-channel attack mitigation. It also enables cross-domain authentication and multi-vehicle authentication across diverse scenarios. We conducted a comprehensive security deployment analysis by employing widely recognized Real-Or-Random model, along with the ProVerif tool. Furthermore, it exhibits competitive performance metrics in terms of computational efficiency, communication overhead, energy consumption, and average packet loss rate, which demonstrate the potential usability of BCEDAS in various environments.

Cross-modal image-text retrieval has garnered significant attention for its ability to integrate visual and textual modalities, thereby enhancing graphic-text interpretation and cross-modal reasoning. This integration addresses key challenges such as semantic heterogeneity, low retrieval efficiency, and generative bias. However, most existing approaches prioritize retrieval accuracy through the use of complex neural architectures but pay insufficient attention to privacy preservation, thereby limiting their suitability for deployment in resource-constrained IoT environments. To address these limitations, we propose LPPITR, a novel lightweight and privacy-preserving cross-modal retrieval framework. The scheme utilizes knowledge distillation to train compact student models for feature extraction from both images and texts, mapping them into a shared semantic space. To improve retrieval performance, we introduce a novel training objective that combines cross-modal triplet ranking loss with intra-modal weighted contrastive loss, thereby enhancing hard-negative discrimination. Moreover, built upon the DT-PKC cryptosystem, we construct an encrypted adaptive overlapping index tree integrating FCM clustering and a Cumulative Confidence Coverage strategy. Combined with a dynamic backtracking mechanism, this structure significantly reduces ciphertext retrieval latency and guarantees exact search precision. Extensive experiments on the COCO 5K and Flickr30K datasets show that LPPITR delivers strong retrieval accuracy and high efficiency, validating its effectiveness and practicality.

The underwater Internet of things (UIoT) has become a prominent technology in next-generation ocean monitoring systems, with autonomous underwater vehicles (AUVs) serving as critical mobile edge sensing platforms for sensor network data collection. The deployment and application of 5 G technology in UIoT mark a significant advancement in underwater network communication capabilities. However, the performance of UIoT is critically constrained by two fundamental challenges: the dynamic and unpredictable nature of underwater environments, and the energy limitations of AUV serving as mobile edge nodes. For instance, node mobility caused network instability, affecting data collection efficiency. And high and uneven energy consumption leads to shortened network lifetime. Moreover, limited AUV energy results in AUV loss and diminished data collection efficiency. To solve these problems, an energy-efficient optimization data collection algorithm based on mobile edge sensing in 5 G UIoT (EEODC-MES) is proposed in this paper. In EEODC-MES, the network clustering is constructed by analyzing the movement characteristics of sensor nodes, and a cluster-head node is selected. Subsequently, the reward for edge sensing device (AUV) collecting data from cluster-head nodes is calculated based on the payoff matrix. The cluster-head node with the highest reward value is prioritized for AUV visitation. AUV decides whether to continue visiting cluster nodes or return to base based on its remaining energy and return energy consumption. The performance of EEODC-MES is compared with that of other data collection algorithms, namely greedy and adaptive AUV path-finding (GAAP), AUV-aided energy-efficient data collection (AEEDCO), and traveling salesperson problem (TSP). Compared with GAAP, AEEDCO, and TSP, EEODC-MES, respectively, improves the network lifetime by 31.8%, 30.1%, and 7.1%. Compared with GAAP and TSP, EEODC-MES, respectively, reduces the collection delay by 26.08% and 51.77%.

Face recognition system is vulnerable to the threat of adversarial attacks. By adding small perturbations to the original image, attackers generate adversarial samples to mislead the target model, highlighting the importance of research on face recognition against aggression. However, existing face recognition attack methods are not satisfactory due to the low transferability and the lack of imperceptibility. In this paper, we propose a new attack framework: Saliency Semantic Ranking with Quality Restoration Synergistic Adversarial Attack on Face Recognition (SRRA). The framework divides the face into semantic regions, sorts them according to the saliency map, and finds the most critical region. In addition, the interpolation-based quality restoration module is designed to dynamically interpolate the original and generated images to improve the imperceptibility of adversarial samples. Extensive experiments on public datasets CelebA-HQ and FFHQ prove that the adversarial samples generated by SRRA have better transferability and imperceptibility compared with the existing methods. This study constructs a new attack architecture for face recognition system, and provides a new technical reference for system security evaluation and design.

Dynamic Searchable Symmetric Encryption (DSSE) extends static encryption by supporting dynamic operations such as encrypted database updates. While the security notion of forward and backward privacy ensures that clients cannot infer newly added files via search operations or deduce deletions through prior query tokens, DSSE’s vulnerability to leakage-abuse attacks remains insufficiently explored. We construct a subgraph attack framework tailored for dynamic searchable encryption environments. The framework leverages file length and retrieval size metadata extracted from access patterns to enable precise inference of query-keyword associations through advanced pattern analysis techniques. Furthermore, we introduce a Jaccard similarity threshold (JST) as a quantitative metric to evaluate the similarity between user queries and document keyword sets. By integrating JST with frequency information extracted from search patterns, our method significantly improves the precision of keyword recovery. Experimental validation in dynamic scenarios demonstrates the efficacy of our approach. On the Enron and Lucene datasets, the proposed attack achieves remarkable accuracy rates exceeding 95

The proliferation of multimodal data in cloud-based applications creates an increasing demand for efficient cross modal retrieval while protecting sensitive information from untrusted servers. Existing privacy-preserving cross-modal retrieval schemes often suffer from limited accuracy or high search overhead. To address these limitations, this paper proposes LMCR, which integrates lossless encrypted scoring with multi probe hybrid indexing. LMCR uses CLIP to map image and text samples into a shared semantic space and further applies Cross-Modal Consistency-Aware Density Refinement (CMDR) to construct semantic partitions jointly supported by image and text neighborhoods. Based on these partitions, semantic inverted lists and cluster-specific KD-Trees are built to reduce the encrypted search space through multi-probe retrieval. In addition, an orthogonal-LWE-based mechanism preserves plaintext ranking during encrypted linear search. Experiments on four benchmark datasets demonstrate that LMCR achieves high retrieval accuracy and substantial search efficiency.

Large language models (LLMs) have demonstrated remarkable potential in instruction-driven recommendation, where tasks and personalization are expressed through natural language prompts. However, their behavior often remains opaque and unstable, limiting practical deployment. To address this, we propose Staged-Reasoning Prompting (SRP)—a structured methodology that decomposes recommendation into multiple well-defined reasoning stages. Each stage employs targeted prompt templates with explicit input–output contracts, transforming the LLM from a black box into a transparent, controllable, and reusable component. SRP’s modular architecture enhances interpretability, debuggability, and efficiency, enabling flexible substitution and tuning of individual stages without redesigning the entire pipeline. We further design several prompt strategies for the key summarization and reasoning stages and conduct comprehensive experiments across four benchmark datasets. Results show that SRP consistently outperforms direct prompting and state-of-the-art baselines, achieving competitive accuracy with a much smaller LLM (e.g., 7B parameters) while substantially reducing inference latency and token cost. Overall, SRP demonstrates a cost-effective and generalizable paradigm for building reliable, LLM-driven software systems through structured prompt engineering.

Dynamic searchable symmetric encryption (DSSE) enables a server to efficiently searches and updates over encrypted files with the price of unintentionally sensitive information leakage that can be exploited by adversaries to infer keywords queried by users. Recently, Xu et al. (USENIX Security 2023) proposed volumetric inference attacks (VIA) based on response volume patterns demonstrate that DSSE schemes with advanced security properties are also vulnerable to Leakage Abuse Attacks (LAAs). However, their attack only utilizes the equational relationship between volume leakage and has a undesirable accuracy with insufficient leakage information. Therefore, in this work, we fully analyze the existence of multi-features of volume pattern in real datasets, such as variance and skewness, and incorporate them to further improve the accuracy, and the results prove that our attack, dubbed VFFIA, is much better than VIA and has strong interference resistance under insufficient auxiliary knowledge. The VIA attack almost fails in the 30

As a privacy-preserving distributed machine learning paradigm, federated learning (FL) has been widely adopted in critical sectors, including healthcare and finance. However, its open participation mechanism exposes the global model to severe threats from backdoor attacks. Existing defense mechanisms are often constrained by the coupling of benign statistical heterogeneity and malicious attack stealthiness within the feature space. To address these challenges, we designed direction alignment and sign consensus backdoor detection (DABD), a novel dual-dimensional consistency detection framework combined with adaptive weighted fusion. In particular, first, it employs cosine similarity for computing a direction alignment score, capturing shifts in the global optimization trend. Then, it utilizes a majority voting mechanism on Top-K parameter updates to derive a sign consensus score, identifying subtle malicious sign flipping on critical dimensions often overlooked by holistic directional metrics. Finally, a temperature-controlled SoftMax function maps these dual scores into aggregation weights, enabling threshold-free suppression of malicious gradients. Extensive experiments on three benchmark datasets demonstrate that DABD effectively mitigates sophisticated attacks, including Neurotoxin, achieving substantial improvements in the main task accuracy and considerable reductions in the attack success rate as compared to the state-of-the-art baselines.

With the widespread application of Internet of Things (IoT) technology in the medical field, real-time collection and transmission of medical images become feasible. However, existing privacy-preserving image retrieval schemes often suffer from low efficiency and high communication overhead when operated in resource-constrained IoT environments due to the lack of efficient models. Thus, achieving efficient and secure medical image retrieval on limited-resource devices has emerged as a critical challenge. To address this, we propose LPPUBR, a lightweight, privacy-preserving, unsupervised bitmap retrieval scheme designed for IoT environments with constrained resources. LPPUBR utilizes a secure and lightweight deep learning model to extract deep feature descriptors from images and employs product quantization (PQ) to encode them into binary bitmaps, enhancing retrieval efficiency while reducing computational and storage costs. Particularly, a cross-quantization contrastive learning strategy is applied to jointly train the neural network model and PQ codewords for unsupervised learning. Furthermore, to improve interaction efficiency and reduce communication costs among multiple servers, we optimize the intermediate value recovery operation and redesign the related protocols in $n$n-party secret sharing using a group communication strategy. A comprehensive theoretical analysis and experimental evaluation demonstrate that LPPUBR maintains retrieval accuracy comparable to the original unsupervised model while ensuring data security. Moreover, LPPUBR surpasses existing schemes in terms of computational cost, communication overhead, and retrieval efficiency.

Mobile Underwater Internet of Things (UIoT) has become one of the pivotal technologies in the advancement of smart oceans monitoring systems, with autonomous underwater vehicles (AUVs) offering a highly promising way for collecting data from mobile sensor nodes. In Mobile UIoT, the mobility of sensor nodes leads to dynamic changes in network topology, making it difficult for the network to form stable clusters. The selection and position of cluster-head nodes will also dynamically change, which increases the difficulty of AUV collecting cluster-head nodes data and further increases data collection delay. To overcome these above difficulties, AUVs are employed as mobile edge devices for collecting data from nodes within the UIoT. In addition, a data collection algorithm orientated for dynamic network topology (DCADNT) is proposed in this paper, which includes three phases: dynamic clustering, dynamic transmission, and dynamic data collection. DCADNT not only considers the impact of dynamic topology on data collection, but also takes into account the impact of propagation delay. By calculating the packet delivery probability based on propagation delay, the reliability of data transmission can be effectively guaranteed. The performance of DCADNT is evaluated by extensive simulations and the compared results with other typical data collection algorithms are given, which demonstrate that DCADNT effectively reduces data collection delay and improves packet collection rate.

The growing severity of software vulnerabilities has made efficient vulnerability detection increasingly critical. Large language models (LLMs), with their powerful natural language processing capabilities, offer new research potential for this task. However, existing prompt-based vulnerability detection methods typically rely solely on source code text and fail to fully exploit the deep semantics and structural logic of the code. To address this issue, this paper proposes MAIAP, a vulnerability detection method based on multi-dimensional auxiliary information–enhanced prompting. MAIAP structurally integrates essential information from traditional program analysis—API call sequences, data flow dependencies, and code property graphs—into the prompt, enabling LLMs to perceive program logic flows during reasoning. Furthermore, by leveraging the multi-turn contextual memory capability of the DeepSeek model, we design an enhanced prompt framework tailored for vulnerability detection, thereby improving contextual understanding and classification accuracy. Experimental results on Java and C/C++ vulnerability datasets show that MAIAP achieves accuracies of 90.6% and 91.8%, outperforming the state-of-the-art baselines by 16.3% and 18.4%, respectively. We further validate the effectiveness of various auxiliary information types in improving both performance and interpretability.

With the rise of cloud computing, large-scale image data from personal, medical, and enterprise archives are often outsourced to cloud servers for efficient storage and computation. To ensure privacy, sensitive images must be encrypted before uploading. Cloud service providers (CSP) offer Database as a Service (DBaaS), including secure image retrieval, for managing encrypted data. Many existing schemes aim to enable privacy-preserving image retrieval but face challenges such as low retrieval efficiency, high computational costs, and limited access control. This paper presents an Efficient and Controllable Privacy-Preserving Image Retrieval (ECPIR) scheme for scenarios using cloud-based database services. We design a hierarchical graph index to organize image vectors in multi-level formats, improving retrieval efficiency. Additionally, we propose a lightweight polynomial-based access control strategy, FastPolyAccess, which uses Fast Fourier Transform (FFT) to enhance computational efficiency and manage access for large-scale user bases. Experimental results show that ECPIR offers superior retrieval performance and robust access control while ensuring privacy.

Federated learning was shown to be vulnerable to backdoor attacks, while existing defense methods were mainly dependent on single statistical features in the parameter space. Under non-independent and identically distributed (Non-IID) conditions, the overlap between benign and malicious update features often led to degraded detection performance, and dedicated defense schemes for low-rank fine-tuning scenarios were still insufficiently investigated. To address these issues, a backdoor defense scheme for federated learning based on dual-domain synergistic detection was proposed. In the frequency domain, discrete cosine transform (DCT) was used to decouple client updates, and semantic alignment together with spectral density clustering was introduced to improve the separability between benign and malicious updates. In the parameter domain, Low-Rank Adaptation (LoRA) spectral consistency detection was adopted to identify anomalous deviations in the low-rank subspace, and causal probing was integrated to actively expose hidden backdoors. Experimental results showed that the proposed scheme effectively decreased the attack success rate of multiple backdoor attacks on both the CIFAR-10 vision task and the Llama-2-7B-based federated LoRA fine-tuning task. Fed-DCR provides a favorable balance between backdoor mitigation and main-task performance in most experimental settings.