In recent years, deep learning models have made significant advancements in enhancing robustness against single-perturbation adversarial attacks, such as$\ell_{p}$-norm attacks. However, the development of defense mechanisms for composite attacks involving multiple semantic perturbations remains a challenge. In this paper, we propose a method that combines projected gradient descent (PGD) with sequential semantic perturbations to generate composite adversarial examples (CAEs), providing a more comprehensive evaluation of model robustness in various scenarios. Existing adversarial training methods primarily focus on improving robustness against single-type attacks. We introduce learnable adversarial training (LAT), which leverages the classification boundaries of clean models to guide the training of robust models. In contrast, our approach not only enhances the model's defense against composite perturbations but also significantly reduces the loss of natural accuracy. Experiments on the CIFAR-10, CIFAR-100 and Tiny ImageNet datasets show that our proposed training method outperforms traditional$\ell_{\infty}$-norm boundary-based adversarial training, demonstrating superior robustness against various attack types. In summary, our method strikes an optimal balance between defense against composite perturbations and natural accuracy, showing strong potential for practical applications.

Document-level Event Extraction (DEE), a core task in natural language processing (NLP), aims to automatically identify and extract event information from documents. Despite significant progress in this area, two challenges remain: (1) existing approaches often overlook the guidance provided by event schemas during DEE; (2) they fail to simultaneously account for local and global interactions among dispersed argument mentions. To address these issues, we propose a novel DEE framework that leverages event schemas and knowledge graphs as guiding priors. Specifically, domain-specific event schemas are manually extracted and organized into a knowledge graph, serving as prior knowledge to guide the extraction process. Additionally, we construct a document knowledge graph based on positional and semantic relations among sentences and entity mentions, employing representation learning to capture complex inter-mention interactions effectively. Experiments conducted on the publicly available financial benchmark dataset ChFinAnn demonstrate that our method achieves significant improvements in F1 scores compared to state-of-the-art (SOTA) models, validating its effectiveness and advancement.

Adversarial training (AT) is a fundamental approach for enhancing the robustness of deep neural networks. However, its main limitation lies in the substantial computational cost required to generate adversarial examples. To mitigate this issue, researchers have proposed single-step adversarial training methods. However, methods such as the Fast Gradient Sign Method (FGSM) are prone to catastrophic overfitting (CO), which leads to a sharp decline in model robustness during training, limiting their practical applicability. To this end, this paper proposes an improved algorithm (FGSM-AWA) based on an adaptive fusion mechanism, which balances the contribution of adversarial examples to model training by dynamically adjusting the fusion factor. Additionally, the algorithm incorporates a specialized loss function designed to combat CO, ensuring that models affected by overfitting are restored to a stable training state. We validate our approach through experiments on CIFAR-10, CIFAR-100 and Tiny ImageNet datasets, employing three network architectures: ResNet-18, WideResNet-34 and PreActResNet-18. Experimental results demonstrate that FGSM-AWA outperforms baseline methods in both robustness and training efficiency.

In recent years, deep learning neural networks have made remarkable progress, especially in the application of computer vision. Although these networks have extraordinary capabilities, they are obviously vulnerable when they are subjected to malicious interference. Although researchers have proposed many methods to improve the robustness of the model to interference, the existing antagonistic training methods still have obvious limitations: excessive consumption of computing resources and disastrous catastrophic overfitting. In order to meet these challenges, we put forward Dynamic Label Smoothing (DLS) and Differential Disturbance Strategy (DPS), and developed a new countermeasure training framework, FGSM-DSEM, which integrates dynamic exponential moving average and momentum-guided gradient direction. A large number of experiments on different network architectures and data sets show that the FGSM-DSEM method significantly enhances the robustness of the model, while significantly reducing the risk of the catastrophic overfitting problem, and achieves these improvements with minimum training overhead.

In recent years, satellite networking technology has garnered significant attention and found applications across various domains, including edge computing, data communication, and data storage. Satellite networks offer a direct solution to network coverage issues, leveraging advantages such as extensive coverage, minimal reliance on infrastructure, and independence from cost and distance constraints. Despite the widespread use of erasure codes for fault tolerance in distributed storage systems, current solutions often lack direct support for the dynamic topologies characteristic of satellite networks. To address the data transmission and latency challenges faced by erasure code systems in dynamic satellite networks, we propose a novel index, i.e., Era-H2H index, that integrates space-time graphs with tree decomposition. Additionally, to reduce the storage costs associated with Era-H2H index, we present an optimization algorithm that leverages the recovery properties of erasure codes and the temporal features of the space-time graph. To further minimize recovery costs, we develop a two-layer regeneration tree, i.e., H2H-RegTree, construction algorithm based on Era-H2H index. At last, we analyze the complexity of the proposed algorithms and conduct an experimental evaluation to demonstrate their performance.

Deep neural network models perform well on image classification tasks, but they are highly susceptible to adversarial samples that add tiny perturbations and output incorrect classification results. The existence of adversarial samples seriously threatens the security of the model itself, in order to mitigate the threat posed by adversarial samples, researchers and scholars have proposed a number of adversarial sample defense methods, which are currently considered to be one of the most effective adversarial defense methods is adversarial training. we redefine the loss in standard adversarial training as a new loss in order to enable the model to better learn the interclass relationships of the examples, which consists of three items: natural categorization loss, target-related loss, and target-irrelevant loss. Through experimental analysis, the target-irrelevant loss has a greater impact on the robustness of the model, and the combination of the three can better improve the robustness of the model, and the DRAT algorithm is further proposed. The experimental results show that the method can further improve the adversarial robustness of the model, and the robust accuracy under multiple attacks outperforms other classical adversarial training methods.

With the wide application of deep neural networks, the security problem of the model is becoming more prominent, adversarial attack is an important tool for evaluating the robustness and security of the model, adversarial attack can be categorized into white-box and black-box attacks. Aiming at the problem of huge perturbation and the low success rate of the adversarial example created in the transfer attack in the black box, a local region approach for randomly segmented channels is given. By randomly segmenting individual dimensions in order to improve the transferability of the adversarial example, the ScoreCAM method is introduced to extract localized focus regions of the image to generate the adversarial example. Experiments show that the performance of the method is better than the baseline algorithm; the fooling rate improved by up to 17%; and the average 2-norm module length decreased by 54.2%.

With the advances in positioning techniques, trajectories are emerged with semantic information such as location-based activities and sign-ins. Essential for applications such as trip recommendations, range queries on semantic trajectories are utilized to identify trajectories that satisfy specific spatio-temporal conditions as well as keyword-related criteria. To reduce the costs of query processing services, data owners often outsource the data services to public platforms, of which the fog-based cloud can offer enhanced storage and computing capabilities and support efficient access to user data. However, outsourcing data in plaintext may raise privacy concerns. To this end, we study the problem of secure range queries on semantic trajectories on the fog-based cloud platform. Initially, to effectively organize semantic trajectories, we design a secure index structure within fog-based cloud framework by distributing spatio-temporal data of trajectories across the fog servers based on semantic contents. Subsequently, we propose a secure range query scheme where keyword pruning prioritizes the fog servers containing query keywords, and then secure spatio-temporal pruning is performed in parallel to obtain the candidate trajectories satisfying spatio-temporal constraints. Finally, on cloud servers, secure timespan verification is applied to ensure that the final results include all query keywords within the specified time span. A comprehensive analysis of the proposed scheme is provided in terms of computational complexity and security guarantees. Leveraging the keyword-first-pruning strategy and parallel processing on fog servers, the query scheme is evaluated to show stable and efficient performance through extensive experiments on real datasets.

Real-time bidding (RTB) is a key component in digital display advertising. During the auction process, advertisers strive to maximize the total value of their winning impressions within the limited budget constraint. However, due to the complexity and volatility of the bidding environment, it is difficult to obtain an optimal bidding strategy. To solve this challenge, the advertising campaign cycle is divided into different time slots to better control and manage advertising delivery. Subsequently, the dual-delay deep deterministic policy gradient (TD3) algorithm is employed to learn the optimal bidding factor for each time slot. This approach allows for the dynamic adjustment of bidding strategies at different time slots, adapting to market fluctuations and target audience behavioral patterns. The analysis shows that direct rewards from the bidding environment were misleading. So the new reward function based on the generalized second pricing (GSP) mechanism, designed to learn optimal policy effectively. Finally, experiments on RTB dataset show that the algorithm could achieve more rewards in different bidding environments.

Direction-aware why-not spatial keyword top-k queries, which aim to retrieve a refined spatial keyword query that includes the missing objects with smallest cost on search direction, have recently received considerable attention from the database community. To offload the data management, when the why-not spatial keyword query processing services are motivated to be outsourced to the cloud for cost savings and flexibility, it may cause serious privacy concerns. To this end, in this paper, we first define and address the problem of privacy-preserving direction-aware why-not spatial keyword top-k queries. To support direction computations in ciphertext, we first propose two novel secure computation protocols, i.e., Secure Modulus Protocol and Secure Angle Computation Protocol. Based on the proposed protocols, to facilitate the secure why-not query processing, we further present a Secure Direction-Aware Why-Not Spatial Keyword Top-k Query (SDAWNkQ) approach for retrieving the refined queries with the smallest penalty by obtaining the best approximate refined directions. Thorough analysis shows the security and computational complexity of our approach, and extensive experimental results on both real and synthetic datasets further demonstrate the query performance of our approach.

In recent years, blockchain technology has received extensive attention and applied in various fields including health-care, IoT and database systems. Utilizing the decentralization and anti-tampering properties, the blockchain provides potential solutions to achieve verification of data queries, without the assumption of the trusted third parties in traditional data verification studies. However, for kNN queries, a common query type in practical location-based scenarios, few existing solutions can directly support the location-based kNN query processing and result authentication based on blockchain. To address this problem, in this paper, we propose a blockchain-based verifiable kNN query processing method. In this method, we first design a novel authenticated data structure called VMR-Tree, which stores the data objects and their neighboring objects in leaf nodes and stores the hash values used for data verification in non-leaf nodes. To verify the results with the minimum size of verification objects (VOs), we design a query result verification method based on the blockchain, in which the client can verify the query results by processing the VOs generated based on the proposed VMR-Tree index and the blockchain. Besides, we further propose an optimization algorithm to reduce the size of VOs. We theoretically analyze the computational complexity and security guarantees of the proposed approaches. We also conducted extensive experiments on real and synthetic datasets to evaluate the efficiency of the proposed method on the result verification of location-based kNN queries.

Attracting research interests and applications in academic and industrial community due to the proliferation of mobile devices, computation and processing services on spatiotemporal trajectory data has usually been outsourced to cloud platforms to save the costs of data storage, computation and management. To prevent the privacy leakage of trajectories from the direct data outsourcing, in this paper, we study the secure similarity search problem on spatio-temporal trajectories and present a secure synchronized spatio-temporal trajectory similarity search approach. In the approach, adopting the Matching Point-point distance Similarity (MPS) measurement, we first propose a Secure Matching Point-point Distance Similarity Computation (SMPSC) Protocol to support the secure similarity calculation on encrypted trajectories. To improve the computational performance, we further propose a Secure Grid Filtering (SGF) method by matching spatio-temporal grid codes to filter the dissimilar trajectories based on the distance threshold in MPS. At last, with SMPSC protocol and SGF method, we propose a Secure Synchronized Spatio-Temporal Trajectory similarity Search Processing (S3TS) method to retrieve similar trajectories based on MPS measurement. We theoretically analyze the computational complexity and security guarantees of the presented approach, and conduct extensive experiments on real and synthetic datasets to demonstrate its search performance.

With the development of location-based services, spatio-textual similarity join has attracted much research attention from academic and industrial communities following the study of spatio-textual data processing. To offload the computation and storage burden of spatio-textual similarity join, outsourcing the data processing and storage to the public cloud can achieve great cost savings, however, may cause serious privacy concerns. To this end, in this paper, we first define and solve the privacy-preserving spatio-textual similarity join problem and propose two novel secure similarity join schemes. As the baseline, we first present a straightforward scheme applying Asymmetric Inner Product Encryption (AIPE) to facilitate the data encryption and similarity calculation in ciphertext. To improve the efficiency of the basic scheme, we further propose an optimized secure top-k spatio-textual similarity join scheme by constructing a secure index based on the hybrid Locality-Sensitive Hashing (LSH). Through matching the encrypted hash values over the secure index to narrow down the candidates, the similarity join results can be efficiently retrieved from the candidate pairs. Comprehensive analysis of the proposed schemes is provided in terms of computational complexity and security guarantees, and extensive experimental results on real and synthetic datasets show the performance of our schemes.

The rapid development of 6G can help to bring autonomous driving closed to the reality. Drivers and passengers will have more time for work and leisure spending in the vehicles, further generating a lot of data requirements. However, edge resources from small base stations are insufficient to match the wide variety of services of the future vehicular networks. Besides, due to the high-speed nature of the vehicles, users have to switch the connections among different base stations, whereas such way will cause external latency during the data request. Therefore, it is vital to enable the local cache of vehicle users to realize the reliable autonomous driving. In this paper, we consider caching the contents in the local cache, small base station, and edge server. In practice, the request preference of some single users may be different from a whole region. To maximize the efficiency of content cache, we design a strategy that uses reinforcement learning algorithm to optimize cache schemes on different devices. The experimental results demonstrate that our strategy can enhance the cache hit ratio by 10%-20% compared with the well-known counterparts.

Software-defined vehicular network (SDVN) is a promising networking paradigm that can provide intelligent information exchanges by separating network management and data transmission. Although the transmission quality of vehicles can be greatly improved by deploying softwarized networking schemes, critical networking issues such as the timeliness of data packets remain due to the dynamic nature of vehicular networks. It is vital to design efficient networking schemes by deeply considering the characteristics of the network, transportation system, and users, to improve overall network performance. To this end, this paper proposes asocial computing inspired predictiverouting scheme (SPIDER) for SDVNs that has a comprehensive consideration to enable low-latency reliable data exchange under dynamic vehicular networks. As for the link lifetime grounded on the vehicular historical data, we introduce the context feature mining and one-shot prediction method to predict vehicle movements with considering the energy saving. We also involve social computing techniques to find the relay nodes with good data spreading abilities. The extensive experiments prove our proposed scheme outperforms four existing schemes.

Adversarial attack is an important research direction of neural network security, and the black-box attack in the unknown model is an important application scenario of adversarial attack. The adversarial samples generated by existing black-box attack algorithms often have drastic hue changes, are easy to be perceived by human eyes, and the algorithm computational efficiency is generally low. Therefore, this paper proposes a multi-scale grid search for the image gray attack strategy, performing perturbations of different intensity on image areas of different scales, and then makes the generated sample hue change more natural. Experiments show that the performance of the method is better than the baseline algorithm, and greatly improves the attack efficiency, increasing the neural network query number over 50% comparing with the fastest baseline algorithm.

In recent years, the computation and processing of trajectories have attracted both academic and industrial communities to study and develop techniques and applications due to the popularization of mobile devices. To offload the data management, it is motivated to outsource the trajectory data to the cloud for achieving great cost savings and flexibility. However, directly outsourcing trajectory data may arise serious privacy concerns. To address the problem, we define the problem of the secure trajectory similarity search over encrypted trajectory data and propose a secure trajectory similarity search approach based on a bi-directional similarity measurement. In this approach, we first propose a secure squared point to line-segment distance computation protocol to facilitate the precious and secure trajectory similarity computations. Furthermore, to improve the search performance, we propose a secure trajectory filtering method based on signature matching to filter out the dissimilar trajectories before the expensive computations on the ciphertext. Based on the proposed protocols and methods, we present a signature-based secure trajectory similarity search processing algorithm to retrieve the similarity results without revealing information about the trajectories. Finally, we theoretically analyze the computational complexity and security guarantees of the proposed approach and conduct extensive experiments on both real and synthetic datasets to test the search performance under various experimental settings.

Outsourcing data services to the public cloud will help data owners save administrative costs, while it may bring privacy concerns. To guarantee the confidentiality of sensitive data against the untrusted cloud and unauthorized users, one way is to have the cloud perform the query processing with encrypted query requests over outsourced encrypted data. However, it is of great challenge to support various types of query processing, especially those of answering why-not questions, under data encryption settings without sensitive information leakage. In this paper, we define and solve the problem of answering why-not questions on top-$k$queries with privacy protection. To address the problem, we propose two Secure Why-Not top$k$Query (SWN$k$Q) processing approaches. In the basic approach, we adopt the Paillier cryptosystem to guarantee the semantic security of data and queries and propose two new secure protocols to support computations on ciphertext in SWN$k$Q processing. Then, we propose a secure weighing space generation method for obtaining the best approximate refined query. To solve the efficiency problem of the basic approach, we further propose an optimized approach, in which dominance-based secure data pruning and early stopping conditions are presented to improve the query efficiency by pruning searching space in the retrieval of refined queries. Thorough analysis shows the security and computational complexity of our approaches, and extensive experimental results on real datasets further demonstrate the query performance of our approaches.