Abstract Model stealing attacks in the Machine Learning as a Service (MLaaS) paradigm face multiple technical challenges, particularly in scenarios lacking training data support. Target models remain vulnerable to reverse engineering and replication under strict environmental constraints. Existing data-free model stealing methods face critical issues such as excessively high query costs, limited attack accuracy, and low sample utilization efficiency. These factors significantly undermine the practical feasibility of such attacks. In this work, we design a Multi-Generator Collaborative Training Approach for Data-Free Model Stealing (MGCT) to address these technical bottlenecks. This framework employs a two-phase optimization paradigm to achieve effective extraction of the target model: In the first phase, a parallelized generator architecture is utilized for multi-generator collaborative training, which enhances the diversity of synthetic samples and alleviates the optimization difficulty of a single generator through a load-balancing strategy. A model distillation mechanism based on an enhanced experience replay is designed in the second phase. By integrating a hybrid sampling strategy and class-balancing constraints, low-confidence samples are dynamically selected, and class distribution equilibrium is maintained, thereby significantly improving the representational capacity of the training data. Compared to existing baseline methods, experimental evaluation results demonstrate that the MGCT framework significantly improves core evaluation metrics, including clone accuracy and query efficiency. Specifically, experimental data on the CIFAR-100 dataset show that the MGCT framework improves cloning accuracy by 2.07% and achieves 14.14% improvement under constrained budgets. These experimental results demonstrate the efficacy of the proposed method in addressing the technical bottlenecks of data-free model stealing.

Abstract Recent studies have indicated that large language models (LLMs) remain susceptible to adversarial attacks, despite enhanced robustness through the chain-of-thought (CoT) capability. However, this capability also introduces the potential for more covert and effective adversarial attack methods. This paper proposes a CoT Transfer Adversarial attack framework (CTTA) for general LLMs. Initially, we utilize a pre-trained model based on the transformer architecture and fine-tune it on various tasks to serve as a surrogate model. Subsequently, different levels of adversarial attack algorithms are utilized, and the generated adversarial samples are used as transfer samples. A thought chain-based adversarial transfer attack framework is constructed using transfer samples and thought chain techniques. Finally, various indicators are utilized to assess the performance of the general LLMs in response to this attack. The results demonstrate that the attack framework surpasses current state-of-the-art research. Numerous experiments on LLMs with varying performance and parameter sizes have validated the effectiveness, stability, and generalizability of this attack. The model’s error response and the superiority of this attack are thoroughly examined using attention by gradient technology, confirming the security threats posed by LLMs when leveraging CoT capability. This has significant implications for enhancing the security and robustness of LLMs.

Abstract The primary challenge of the multimodal sentiment analysis (MSA) task is the modal fusion, and the lack of modalities may exist in the fusion process, which leads to poor prediction results. Most of the previous research on multimodal fusion is single‐stage fusion, disregarding how various modality subsets interact, as well as rarely considering relative position relationship of modality sequences, causing the fragmentation of context info. Considering the aforementioned issues, this study introduces an MSA method based on the multi‐stage graph fusion network (MSGFN) under random missing modality conditions to improve the robustness of the model to MSA under the random missing modality conditions. Firstly, for each modality, its inter‐modal and intra‐modal multi‐head attention are used to learn robust representation of the modality sequence. Meanwhile, the relative position encoding (RPE) is introduced into mechanism of attention that enables model to perceive and learn the relative position before and after the modality sequence when calculating attention, thereby better understanding the contextual info of the sequence. After that, the transformer encoder receives the learned modality features and uses the pre‐trained model to supervise the reconstruction of the missing modality information. Finally, the feature representations of different modalities have effectively fused using multi‐stage graph fusion network, and the output is used for the ultimate sentiment classification. Wide experiments are conducted on two publicly available datasets, CMU‐MOSI and IEMOCAP, and the findings indicate that the proposed method can better handle the challenges caused by modality fusion and modality missing compared with several baseline methods.

GIFT, as a lightweight block cipher algorithm, is mainly suitable for resource-constrained environments, such as the Internet of Things, to achieve efficient cryptographic communication. For the diffusion characteristics of round function of the GIFT algorithm, this article proposes two byte-based differential fault attack models. Theoretically, the first and second models require 53.44 and 12.42 byte faults to recover the master key. Experimental results show that the first and second models require about 79 and 16 byte faults, respectively, to recover the master key. The models proposed in this article have made significant breakthroughs in terms of both the attack range and the number of required faults, which provide important guidance for algorithmic security research and the design of fault-tolerant mechanisms for IoT devices.

AEGIS is a lightweight cipher for resource constrained environments such as the Internet of Things (IoT), which is one of the seven algorithms that ultimately won the CAESAR competition, and has become a focus of cryptographic research. Differential fault attack is a method to recover important secret data by researching the difference in state before and after injecting a fault, which is one of the most serious threats to lightweight cryptographic algorithms. This paper propose differential fault attacks on AEGIS based on the random bit fault model, which theoretically respectively requires only 141, 176, and 246 random bit faults to fully recover the intermediate state of AEGIS-128, AEGIS-256, and AEGIS-L, which are far superior to the known results in terms of the number of faults, computational complexity, and difficulty of models. The attack method on AEGIS contributes to assessing the security of IoT devices and systems, while also guiding the design of IoT security defense strategies. The approach proactively detects vulnerabilities in lightweight cryptographic algorithms implemented on resource-constrained devices such as sensors and smart home chips, thus driving enhancements in algorithms and hardware, and has great significance for the security of IoT platforms.

The tight cyber–physical coupling in industrial control system (ICS) makes it vulnerable to attacks, where attackers can exploit system vulnerabilities to cross domain boundaries, causing significant losses. A novel quantitative risk assessment framework for ICS is proposed, integrating both the cyber domain and the physical domain to address potential risk assessment issues. First, an entropy-weighted technique for order preference by similarity to ideal solution method introducing triangular fuzzy numbers is proposed to solve the vulnerability index for multiattribute decision making to obtain the a prior probability. Second, the risk propagation mechanism of complex network topology and device interaction under ICS was studied, and the posterior probability model based on susceptible-exposed-infectious-recovered was designed to dynamically capture the propagation characteristics of risks in the network. Finally, the safety loss level is defined based on standard criteria to achieve quantitative risk assessment of ICS. The proposed risk assessment model was experimentally validated on the SWaT testbed, with results confirming its feasibility and effectiveness.

In recent years, public security incidents caused by deepfake technology have occurred frequently around the world, which makes an efficient and accurate deepfake detection model crucial. The existing advanced methods use the manipulation features in the image to realize the binary classification of real and fake images by training complex neural network models. However, these models rely on a single manipulation feature, and the detection accuracy of these methods will be greatly reduced when the forgery technology or image quality of the training dataset and the validation dataset are different. Inspired by the existing work, we propose a two-stream collaborative learning framework that combines spatial texture differences and frequency information. The average difference convolution (ADC) is designed to extract the spatial texture difference information of the image, and the gray image frequency-aware decomposition (GFAD) is used to extract the artifact information of the image in the frequency domain. At the same time, the ViT idea is combined with cross-attention mechanism for feature fusion to comprehensively mine forged features in forged images. Experimental results show that the proposed model has good detection effects on three benchmark datasets. In terms of cross-dataset evaluation, the AUC on Celeb-DF dataset reaches 82.86%, which is better than the existing advanced methods.

Indiscriminate data poisoning attacks are highly effective against unsupervised learning. However, recent studies show that contrastive learning is also susceptible to data poisoning attacks. As a form of data poisoning attack, the attacker adds poison to the clean pre-training dataset. This article proposes IDPA, an indiscriminate data poisoning attack targeting the encoder in contrastive learning. where the attacker’s goal is to directly poison the pre-trained encoder. The feature vectors of any clean sample and the attacked sample from the attacker will exhibit high similarity, causing the downstream classifier to misclassify the clean sample as the samples designated by the attacker. Therefore, this article formulates IDPA as a dual optimization problem and defines two loss functions: the attack effectiveness loss and the model utility loss. These losses are associated with effectively poisoning the pre-trained encoder and maintaining the accuracy of the downstream classifier, respectively. During training, the attack affects the contrastive learning algorithm and predictions are made on multiple datasets. Experimental results show that the attack success rate of 92%. This article evaluates the effectiveness of IDPA on the CLIP dataset released by OpenAI, with attack success rate of 88%.

Abstract Image inpainting aims to repair the damaged region according to the known content in the damaged image. Recently, image inpainting methods have poor effects on high‐resolution damaged images, and the research on the inpainting of large‐area damaged images is limited. Therefore, this paper proposes an image inpainting method based on Generative Adversarial Networks (GAN) inversion and autoencoder. This work consists of two phases: first, the authors design an autoencoder‐based GAN, which learns the mapping from noise to low‐dimensional feature maps by training a generator, and then converts the generated feature maps into high‐resolution images. Thus, the difficulty of learning the mapping relationship is reduced. Second, the authors adopt the learning‐based GAN inversion to infer the closest latent code. The trained GAN is then used to reconstruct the complete image. Finally, the authors compare their method with other classical methods on the CelebAMask‐HQ, Flickr‐Faces‐HQ, and ImageNet datasets. According to the quantitative comparison, when the mask range is large, in other words, when the image has a large area of damage, the authors’ method is superior to the comparison methods. According to the qualitative comparison, the structure of the high‐resolution image inpainted by the authors’ method is more reasonable and the texture details are more realistic.

Aiming at the problem of adopting the same level of privacy protection for sensitive data in the process of data collection and ignoring the difference in privacy protection requirements, the authors propose an adaptive personalized randomized response method based on local differential privacy (LDP-APRR). LDP-APRR determines the sensitive level through the user scoring strategy, introduces the concept of sensitive weights for adaptive allocation of privacy budget, and realizes the personalized privacy protection of sensitive attributes and attribute values. To verify the distorted data availability, LDP-APRR is applied to frequent items mining scenarios and compared with mining associations with secrecy konstraints (MASK), and grouping-based randomization for privacy-preserving frequent pattern mining (GR-PPFM). Results show that the LDP-APRR achieves personalized protection of sensitive attributes and attribute values with user participation, and the maxPrivacy and avgPrivacy are improved by 1.2% and 4.3%, respectively, while the availability of distorted data is guaranteed.

Abstract As an auto-parallelization technique with the level of thread on multi-core, Thread-Level Speculation (TLS) which is also called Speculative Multithreading (SpMT), partitions programs into multiple threads and speculatively executes them under conditions of ambiguous data and control dependence. Thread partitioning approach plays a key role to the performance enhancement in TLS. The existing heuristic rules-based approach (HR-based approach) which is an one-size-fits-all strategy, can not guarantee to achieve the satisfied thread partitioning. In this paper, an importancedegree basedthreadpartitioningapproach (IDaTPA) is proposed to realize the partition of irregular programs into multithreads. IDaTPA implements biasing partitioning for every procedure with a machine learning method. It mainly includes: constructing sample set, expression of knowledge, calculation of similarity, prediction model and the partition of the irregular programs is performed by the prediction model. Using IDaTPA, the subprocedures in unseen irregular programs can obtain their satisfied partition. On a generic SpMT processor (called Prophet) to perform the performance evaluation for multithreaded programs, the IDaTPA is evaluated and averagely delivers a speedup of 1.80 upon a 4-core processor. Furthermore, in order to obtain the portability evaluation of IDaTPA, we port IDaTPA to 8-core processor and obtain a speedup of 2.82 on average. Experiment results show that IDaTPA obtains a significant speedup increasement and Olden benchmarks respectively deliver a 5.75% performance improvement on 4-core and a 6.32% performance improvement on 8-core, and SPEC2020 benchmarks obtain a 38.20% performance improvement than the conventional HR-based approach.

Industrial Internet plays an important role in key critical infrastructure sectors and is the target of different security threats and risks. There are limitations in many existing attack detection approaches, such as function redundancy, overfitting, and low efficiency. A combinatorial optimization method—Lagrange multiplier—is designed to optimize the underlying feature screening algorithm. The optimized feature combination is fused with random forest and XG-Boost selected features to improve the accuracy and efficiency of attack feature analysis. Using both the UNSW-NB15 and natural gas pipeline datasets, we evaluate the performance of the proposed method. It is observed that the influence degrees of the different features associated with the attack behavior can result in the binary classification attack detection increasing to 0.93 and the attack detection time reducing by 6.96 times. The overall accuracy of multi-classification attack detection is also observed to improve by 0.11. We also observe that nine key features of attack behavior analysis are essential to the analysis and detection of general attacks targeting the system, and by focusing on these features one could potentially improve the effectiveness and efficiency of real-time critical industrial system security. In this article, the CICDDoS2019 and CICIDS2018 datasets are used to prove the generalization. The experimental results show that the proposed method has good generalization and can be extended to the same type of industrial anomaly datasets.

Industrial internet security incidents occur frequently, and it is very important to accurately and effectively detect industrial internet attacks. In this paper, a novel CNN-LSTM fusion model-based method is proposed to detect malicious behavior under industrial internet security. Firstly, the data distribution is analyzed with the help of kernel density estimation, and the Pearson correlation coefficient is used to select the strong correlation feature as the model input. The one-dimensional convolutional neural network and the long short-term memory network respectively extract the spatial sequence features of the data and then use the softmax function to complete the classification task. In order to verify the effectiveness of the model, it is evaluated on the NSL-KDD dataset and the GAS dataset, and experiments show that the model has a significant performance improvement over a single model. In the detection of industrial network traffic data, the accuracy rate of 97.09% and the recall rate of 90.84% are achieved.

This paper proposes a hierarchical framework for industrial Internet device authentication and trusted access as well as a mechanism for industrial security state perception, and designs a cross-domain authentication scheme for devices on this basis. The scheme obtains hardware device platform configuration register (PCR) values and platform integrity measure through periodic perception, completes device identity identification and integrity measure verification when device accessing and data transmission requesting, ensures secure and trustworthy access and interoperation of devices, and designs a cross-domain authentication model for trustworthy access of devices and related security protocols. Through the security analysis, this scheme has good anti-attack abilities, and it can effectively protect against common replay attacks, impersonation attacks, and man-in-the-middle attacks.

Digital disinformation, such as those occurring on online social networks (OSNs), can influence public opinion, create mistrust and division, and impact decision- and policy-making. In this study, we propose a disinformation diffusion trend analysis and identification method, which uses social situation analytics and a multilevel attention network. First, we present a division and feature representation approach of social user circle based on the content sequence (internal driving factor) and social contextual information (external driving factor) of users associated with disinformation. Second, disinformation content feature, crowd response feature, and time-series feature are represented using embedding layer and bidirectional long short-term memory neural networks (Bi-LSTMs). We also present an attention mechanism model based on multifeature fusion, which can dynamically adjust the weight of each feature. On this foundation, the fused features are fed into the multilayer perceptron to identify the propagation quantity trend. According to the experimental results of real-world OSNs and social situation metadata, we conclude that while disinformation occurs across OSN platforms, the disinformation is more likely to spread widely in the original OSN platform. We also identify four typical disinformation propagation trends based on propagation patterns and propagation peak times. Findings from our experiments demonstrate that our proposed approach accurately identifies and predicts the diffusion trend of disinformation, which can then be used to inform mitigation strategy.

The dissemination of digital disinformation in online social networks (OSNs) has been the subject of extensive research, although many challenges remain, including the analysis and control of disinformation dissemination across different platforms (i.e., cross-platform). In this article, we investigate and analyze the spreading patterns and regularities of disinformation both within a single platform and across platforms. To explore the complex relationship between user propagation desire and behaviour within the same group, a user propagation desire inference model based on propagation characteristics (behaviour characteristics and time characteristics) and a bidirectional backpropagation (B-BP) deep neural network are constructed. Then, to avoid overfitting due to the interaction of users’ propagation behaviour and the correlation among propagation characteristics, a novel adaptive weighted particle swarm optimization evolutionary algorithm is utilized to further optimize the B-BP deep neural network. We design and conduct a series of evaluation experiments on the current global hot topics including but not limited to novel coronavirus-19 pandemic (COVID-19), food safety, medical and health, and environmental protection. By using a real-world social platform and its social situation metadata analysis, the experimental results show that the proposed method not only accurately predicts the level of user propagation desire under multiple behaviour interactions but also facilitates social platform managers in handling disinformation disseminators. Our findings reveal that the intensity of social users’ desires to spread disinformation is related to the topics and groups that users are interested in, while the propagation motivation of social users is not strong under topics that users are not interested in. Our studies also demonstrate that social users with propagation desires tend to utilize their familiar social platforms and local circles for communication, and the behaviour and desire to spread disinformation to the cross-platform are not strong. We posit that these findings can help inform online and, fine-grained governance and mitigation strategies other than “one size fits all” approaches (e.g., “account prohibition and deletion”), and hopefully minimize disinformation dissemination.

The past decade has witnessed the emergence and progress of multimedia social networks (MSNs), which have explosively and tremendously increased to penetrate every corner of our lives, leisure and work. Moreover, mobile Internet and mobile terminals enable users to access to MSNs at anytime, anywhere, on behalf of any identity, including role and group. Therefore, the interaction behaviors between users and MSNs are becoming more comprehensive and complicated. This paper primarily extended and enriched the situation analytics framework for the specific social domain, named as SocialSitu , and further proposed a novel algorithm for users’ intention serialization analysis based on classic Generalized Sequential Pattern (GSP). We leveraged the huge volume of user behaviors records to explore the frequent sequence mode that is necessary to predict user intention. Our experiment selected two general kinds of intentions: playing and sharing of multimedia, which are the most common in MSNs, based on the intention serialization algorithm under different minimum support threshold ( Min_Support ). By using the users’ microscopic behaviors analysis on intentions, we found that the optimal behavior patterns of each user under the Min_Support , and a user's behavior patterns are different due to his/her identity variations in a large volume of sessions data.

Journal Article Security, Trust and Risk in Multimedia Social Networks Get access Zhiyong Zhang Zhiyong Zhang * Department of Computer Science, Henan University of Science and Technology, Luoyang 471023, P. R. China *Corresponding author: [email protected] Search for other works by this author on: Oxford Academic Google Scholar The Computer Journal, Volume 58, Issue 4, April 2015, Pages 515–517, https://doi.org/10.1093/comjnl/bxu151 Published: 23 December 2014 Article history Received: 26 November 2014 Revision received: 28 November 2014 Accepted: 29 November 2014 Published: 23 December 2014

Multimedia social networks (MSNs) services and tools provide a convenient platform for users to share multimedia contents, such as electronic book, digital image, audio and video, with each other. However, in an open network, uncontrolled sharing and transmission mode of digital content between users create considerable problems regarding digital rights management (DRM). This paper aims to explore potential paths on the propagation of copyrighted contents. An approach to mining credible potential paths is proposed for MSNs. The formal descriptions were primarily based on rough set theory for mining potential paths. Trust was also measured to find credible potential paths. We presented related algorithms for mining two kinds of paths between any two nodes. Finally, we conducted an experiment based on three non-overlapped sharing communities multiplied by 150 nodes. In the communities found by using a representative real-world MSN YouTube dataset, we further mine the general and credible potential paths based on the simulated trust assessment values. The proposed method could effectively and accurately mine two kinds of potential paths of copyrighted digital content distribution and sharing, which can help to resolve critical DRM issues.

The purpose of delegation is improving resource share and realizing collaboration in CSCW environment. Nowadays agent-based collaboration in multi-agent system lacks of delegation model including formalization and relative mechanisms, especially for variable dynamic collaboration scene. Based on a new proposed concept agent collaboration scene (abbr. ACS), this paper specifies a dynamic agent-based capability delegation model supporting temporal characters and constraint rules, called by capability delegation model for multi-agent system (abbr. CDM for MAS), as well as representing hybrid architecture and key protocol functions related to delegation. The model solves the issue of agent's capability oneness and inflexibility, and realizes dynamic capabilities delegation between agents in an actual application agent-based text information retrieval system, enhancing agent cooperation efficiency and security