<?xml version="1.0" encoding="utf-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:cf="http://www.microsoft.com/schemas/rss/core/2005">
<channel>
<title cf:type="text"><![CDATA[Editorial Board of Journal of Cyber Security -->Volume 11,Issue 2,2026 Table of Contents]]></title>
<item>
<title><![CDATA[A Robust Watermarking Scheme for Deep Neural Networks based on Machine Unlearning]]></title>
<link><![CDATA[http://jcs.iie.ac.cn/xxaqxben/ch/reader/view_abstract.aspx?file_no=20260201&flag=1]]></link>
<description><![CDATA[In recent years, Deep Neural Networks (DNNs) have achieved significant success in many cutting-edge fields, such as image processing, speech recognition, and natural language processing. These DNN models have brought substantial economic benefits to their developers and teams. However, training DNN models requires extensive data and computational resources, with costs that multiply as model parameters increase. Consequently, a well-trained DNN model holds high value for its owner. Unfortunately, high-value, well-trained DNN models face various security threats, including model theft, misuse, and unauthorized distribution. DNN watermarking has become an essential means of protecting model copyrights. Based on whether the watermark is embedded into the model parameters, DNN watermarking can be classified into static and dynamic watermarks. Due to the need for white-box access during verification, static watermarking is challenging to use in practical applications. Dynamic watermarking, which involves adding validation samples and label mappings to DNN models, is vulnerable to watermark removal attacks. Consequently, existing watermarking methods often lack robustness, posing significant risks and security concerns in real-world deployments. This paper proposes a robust watermarking method for DNNs based on machine unlearning. Unlike existing methods, this approach leverages machine unlearning techniques to eliminate original sample mapping-embedded watermarks, replacing traditional sample-label mapping methods to prevent watermark removal attacks, thereby greatly enhancing watermark robustness. Specifically, this method uses a sample selection technique based on sample similarity to identify samples that need to be forgotten. It then selectively forgets the mapping relationships of certain training samples using a gradient ascent strategy to improve watermark robustness. This method effectively counters multiple watermark removal attacks and demonstrates excellent robustness in experiments conducted on CIFAR-10, CIFAR-100, and TinyImageNet datasets, achieving an average watermark extraction accuracy exceeding 98% in the face of various watermark removal attacks.]]></description>
<pubDate>2026/6/4 19:42:18</pubDate>
<category><![CDATA[]]></category>
<author><![CDATA[REN Jixing, XU Wei, WANG Run, LI Boheng, ZHANG Yuyang, WANG Lina]]></author>
</item>
<item>
<title><![CDATA[Robust Fake Audio Detection Algorithm based on Vision Transformer]]></title>
<link><![CDATA[http://jcs.iie.ac.cn/xxaqxben/ch/reader/view_abstract.aspx?file_no=20260202&flag=1]]></link>
<description><![CDATA[With the rapid development of deep learning and speech synthesis technology, speech deepfake technology has made fake speech realistic in terms of naturalness and emotion, posing a great threat to social security. In order to resist the threats to security and privacy brought by these fake technologies, fake speech detection technology based on deep learning has received great attention from researchers and has achieved good performance, but there are still problems with robustness and poor interpretability. Performance degradation occurs significantly when there is a mismatch between training data and actual detection data, and there is a lack of interpretability as existing detection techniques do not provide analysis of detection results. Addressing the issue of poor performance of existing deepfake speech detection techniques under various data mismatch scenarios, this paper proposes a robust speech deepfake detection scheme based on Vision Transformer, which optimizes the entire detection algorithm from both frontend feature extraction and backend neural network aspects. In terms of feature extraction, this paper introduces a frontend feature extractor based on self-supervised learning, which fine-tunes existing generic pre-trained models using labeled data to learn better intermediate speech representations. For the backend neural network, this paper extends Vision Transformer to deepfake speech detection task, decomposing the original positional encoding into time positional encoding and frequency positional encoding. Leveraging the powerful representation capability of Transformer architecture, better feature representations are learned to capture artifacts in the speech to be detected. Experiments indicate that in various complex data mismatch scenarios, our method reduces the detection EER (Equal Error Rate) by 1% to 20% compared to existing methods, exhibiting improved robustness. Additionally, the paper utilizes the attention mechanism of Transformer model to provide interpretability analysis of the decision-making process of the deepfake speech detection model, thus possessing significant practical value.]]></description>
<pubDate>2026/6/4 19:42:18</pubDate>
<category><![CDATA[]]></category>
<author><![CDATA[ZHANG Tong, DENG Junlong, REN Yanzhen, WANG Lina]]></author>
</item>
<item>
<title><![CDATA[A Survey on Textual Backdoor Defense for Language Models]]></title>
<link><![CDATA[http://jcs.iie.ac.cn/xxaqxben/ch/reader/view_abstract.aspx?file_no=20260203&flag=1]]></link>
<description><![CDATA[Language models (LMs) have seen rapid development and are widely deployed across diverse natural language processing (NLP) domains, consistently demonstrating state-of-the-art performance. However, the complex architecture and massive parameter scales of LMs limit their interpretability. Consequently, a range of security threats, particularly backdoor attacks, challenge the reliability and trustworthiness of LMs, impeding their wider deployment. While extensive research aims at defending against backdoor attacks on LMs, most existing methods remain confined to conventional training paradigms, making them ineffective for generative large language models (LLMs). Additionally, current classification standards for textual backdoor defense are inconsistent, and existing reviews either lack comprehensive coverage of the literature or provide insufficient comparative analyses of defenses. To address these gaps and offer valuable insights for future research, this paper systematically reviews and compares a wide range of textual backdoor defenses. Based on the implementation stage and the purpose of the defenders, we categorize the mainstream textual backdoor defense methods into training-stage defense (including trojan weight removal, regularized training, and dataset purifying), and testing-stage defense (including offline model inspection, online input inspection, and regularized decoding). Representative works from each category are subsequently highlighted. Furthermore, this paper summarizes the commonly used datasets and evaluation metrics for textual backdoor defense. By integrating evaluation metrics, we comprehensively analyze the capability requirements of defenders, computational overhead, and defense performance against prevalent textual backdoor attack methods, identifying key limitations of existing defenses. Lastly, we outline future research directions, including developing general defense frameworks, designing tailored defenses for generative LLMs, investigating multilingual defense, exploring the interpretability of textual backdoors, and establishing benchmarks for evaluating backdoor defenses.]]></description>
<pubDate>2026/6/4 19:42:18</pubDate>
<category><![CDATA[]]></category>
<author><![CDATA[WU Zongru, CHENG Pengzhou, ZHANG Zhuosheng, LIU Gongshen]]></author>
</item>
<item>
<title><![CDATA[Quantitative Threat Analysis of Multi-source Security Logs]]></title>
<link><![CDATA[http://jcs.iie.ac.cn/xxaqxben/ch/reader/view_abstract.aspx?file_no=20260204&flag=1]]></link>
<description><![CDATA[Intrusion detection systems (IDS) are critical components of cybersecurity, tasked with identifying and responding to malicious activities. IDS primarily relies on the rules or classification methods to detect anomalies. Rule-based IDSs operate by comparing network traffic against a predefined set of rules to detect anomalies, but they often result in a high false positive rate because they cannot adapt to new scenes. Classification-based IDSs use machine learning algorithms to categorize network traffic as either benign or malicious. These systems often struggle with the granularity required for accurate threat assessment, because the amount of data can overwhelm these systems, leading to important threat indicators being overlooked. To address these limitations, this paper introduces Themis, a novel regression-based framework designed to evaluate and analyze threats present in multi-source security logs. Themis begins by extracting threat entities from web alert logs, which include critical information such as security events and threat IP addresses. These entities are then represented in a multidimensional space, where each dimension corresponds to a specific attribute of the threat entity. To overcome the challenges of data scarcity and class imbalance in security logs, Themis employs unsupervised learning techniques to enhance the features of threat entity samples. The core of Themis is a threat assessment model that leverages these enhanced features to perform threat regression analysis. This model is trained to predict the severity of threats, providing a more precise assessment than traditional intrusion detection methods. To validate the effectiveness of Themis, we conduct detailed regression analysis experiments to explore the dimensions that significantly impact threat severity, as identified through regression analysis. The ablation experiments that demonstrate the benefits of feature-enhanced threat assessment. Furthermore, we compare different regression algorithms used in threat assessment, discussing their respective advantages and disadvantages. Finally, we offer a complexity analysis and practical application recommendations for the various regression algorithms considered.]]></description>
<pubDate>2026/6/4 19:42:18</pubDate>
<category><![CDATA[]]></category>
<author><![CDATA[FENG Wenying, GU Zhaoquan, ZHAO Angxiao, LUO Cui, YUAN Huaping, HU Ning]]></author>
</item>
<item>
<title><![CDATA[Evading Attacks for DeepFake Fake Model Traceability]]></title>
<link><![CDATA[http://jcs.iie.ac.cn/xxaqxben/ch/reader/view_abstract.aspx?file_no=20260205&flag=1]]></link>
<description><![CDATA[In recent years, the proliferation of DeepFakes has caused great alarm among the public and prominent figures. These highly realistic fake images and videos can spread disinformation on a large scale, cause reputational harm, and may even trigger social unrest. In order to deal with the generated fake images and videos, the research in the field of DeepFake forensics has been widely concerned. In the current DeepFake forensics research, DeepFake detection technology is responsible for judging whether a given sample is true or not, while DeepFake traceability technology aims to trace the type of counterfeit model that generates such Deepakes, so as to provide more explanatory results for DeepFake detection. Specifically, DeepFake traceability can be divided into model-schema traceability and model-instance traceability, where model-schema traceability only inferences the specific model schema used, while model-instance traceability attempts to identify model instances with specific training Settings. Both model-architecture and model-instance traceability methods rely on identifying specific traces left by the generation of deepfakes that savvy attackers can destroy or tamper with, rendering the traceability techniques ineffective. It is observed that specific traces used for model traceability exist in both high-frequency and low-frequency components and play different roles in the traceability process. Based on this, this paper proposes an untrained attack evading method—TraceEvader for the first time, and tests it in the most practical non-box setting. Specifically, TraceEvader injects generic imitation traces learned from the original DeepFakes into the high-frequency component and introduces adversarial ambiguity into the low-frequency component to obfuscate the extraction process of certain traces, thereby evading model traceability. In this paper, we experiment with four state-of-the-art model traceability techniques and evaluate their performance in eight generative models, including Generative Adversarial Networks (GANs) and Diffusion Models (DMs) generate representations on forged images. The results show that TraceEvader achieves the highest average attack success rate of 79%, and still shows good robustness in the face of image conversion and professional denoising techniques, and the average attack success rate remains around 75%. TraceEvader confirms the limitations of current model traceability techniques and reminds DeepFakes researchers and practitioners to explore more powerful model traceability techniques.]]></description>
<pubDate>2026/6/4 19:42:18</pubDate>
<category><![CDATA[]]></category>
<author><![CDATA[WU Mengjie, YU Jiayi, WANG Run, YE Xi, ZHANG Yuyang, LIN Chenhao, FANG Liming, WANG Lina]]></author>
</item>
<item>
<title><![CDATA[A Robust Forged Face Detection Scheme Based on Speech-Related Facial Landmarks]]></title>
<link><![CDATA[http://jcs.iie.ac.cn/xxaqxben/ch/reader/view_abstract.aspx?file_no=20260206&flag=1]]></link>
<description><![CDATA[The generation of synthetic audio-visual content is becoming increasingly realistic, posing significant challenges to the detection of falsified video. From the dissemination of fake audiovisuals on social media platforms to misleading content in political propaganda, the potential risks are pervasive. Consequently, the need for effective detection and prevention mechanisms against forged speaker facial videos has become urgent and crucial. However, current mainstream deepfake detection methods struggle to differentiate between compression artifacts and forgery artifacts, leading to a significant drop in detection accuracy in scenarios involving highly compressed videos and social media communications. We propose a Facial-Landmark based Graph Attention Network (FALNet) for detecting forged speaker facial videos, which decouples facial landmarks from video. We introduce a robust video feature extraction network based on facial landmarks and analyze the muscle movements associated with speech behavior, as well as the forged cues introduced during the generation of deepfake speaker facial videos. We designed an adjacency matrix based on facial muscle movements by analyzing the muscle dynamics during speech. The matrix not only preserves the topological information of the face but also effectively captures the differences between genuine and fake facial features. Using a graph attention network as the backbone, we extracted facial features represented by this adjacency matrix. Furthermore, considering the importance of temporal features in video forgery detection, we modeled both short-term and long-term features. Specifically, we first used a graph attention network to capture short-term features and then fed the sequence of short-term features into a recurrent neural network to model long-term dependencies. Experimental results show that our scheme has achieved a detection accuracy of over 98% on video forgery subsets of multiple public datasets. Compared to existing advanced methods based on facial key points, this scheme has achieved a 0.6% to 1.1% improvement in AUC (Area Under the Curve) value. Furthermore, when facing compression, the detection AUC value of this scheme remains above 94%, demonstrating good robustness.]]></description>
<pubDate>2026/6/4 19:42:18</pubDate>
<category><![CDATA[]]></category>
<author><![CDATA[HUANG Yihuan, PENG Li, REN Yanzhen, WANG Lina]]></author>
</item>
<item>
<title><![CDATA[A Survey of Adversarial Techniques Against Large Model Alignment]]></title>
<link><![CDATA[http://jcs.iie.ac.cn/xxaqxben/ch/reader/view_abstract.aspx?file_no=20260207&flag=1]]></link>
<description><![CDATA[With the advent of large models like ChatGPT, the security of AI-generated content has garnered increasing attention from researchers. To ensure that the final behavior of models aligns with human values, alignment techniques play a crucial role during model deployment. These techniques adjust different pre-trained models through fine-tuning or other methods to enhance their reasoning capabilities on specific tasks. Alignment security attacks have attracted widespread attention from academia and industry, but there is currently a lack of systematic review on alignment attack techniques for large models. This paper begins by examining the security risks faced during the deployment stage of aligned large models. It investigates potential vulnerabilities throughout the deployment process and reviews existing alignment techniques. A comprehensive study of current alignment attack methods is conducted, including prompt injection attacks, adversarial attacks, privacy leakage attacks, and backdoor trigger attacks. The analysis identifies security vulnerabilities and potential attack techniques within alignment methods. Secondly, from the perspective of security risks posed by fine-tuning downstream tasks, the paper analyzes how the fine-tuning process compromises the security limitations of aligned large models. It investigates behaviors during the fine-tuning process that may cause alignment security vulnerabilities, providing a detailed analysis of the impact of fine-tuning on the security of secondarily developed models. Thirdly, from the perspective of multimodal development of large models, the paper introduces the architecture of multimodal large language models (MLLM). It summarizes and analyzes the fusion technologies between different modalities within these models and highlights the characteristics of attack concealment due to the continuity of MLLM inputs. Finally, the paper provides an outlook on the future development direction of alignment attack techniques for large models. By deeply exploring the current state and potential risks of alignment attack techniques, the research aims to inspire new ideas and directions in academia.]]></description>
<pubDate>2026/6/4 19:42:18</pubDate>
<category><![CDATA[]]></category>
<author><![CDATA[GONG Runsen, WANG Kai, ZHANG Yulin, ZHANG Weizhe, QIAO Yanchen, ZHANG Yuqing]]></author>
</item>
<item>
<title><![CDATA[Research on Formal Design of SP Network Cipher based on Process Algebra]]></title>
<link><![CDATA[http://jcs.iie.ac.cn/xxaqxben/ch/reader/view_abstract.aspx?file_no=20260208&flag=1]]></link>
<description><![CDATA[In view of the problem that subjectivity and inconsistency in the design and implementation of traditional Substitution-Permutation (SP) network structure cryptographic algorithms, which mainly rely on the designer’s experience and manual implementation, this paper proposes a formal design method based on process algebra to describe the structure of SP network cipher algorithms from the component level. Firstly, the design principles of MCL (MetaCrypto Language) oriented to cryptographic algorithm development are proposed, which lays a foundation for the subsequent formal description. Secondly, through the analysis of the characteristics of SP network structure cryptography algorithm, four components based on process algebra are proposed based on process algebra, and the use of the four components in designing SP network structure ciphers is explained, which is used to formally design and describe SP structure ciphers. Finally, the formal models of TANGRAM cryptography algorithm and SM4 cryptography algorithm are established by using the formal design method, and the key difficulties and technical challenges of formal model design in dealing with complex SP network structure cryptographic algorithms are described. Based on the formal model, MCL model of cipher algorithm is built, which provides a solid theoretical support for the design of block cipher algorithm based on MCL. The correctness is verified by MetaCrypto platform. The verification results show that the design and implementation of lightweight TANGRAM encryption algorithm can be realized based on this method. At the same time, the SP network structure in SM4 cryptographic algorithm can be modeled and implemented correctly. Compared with traditional design methods, formal design methods are superior to traditional design methods in terms of systematicness, accuracy, maintainability and applicability. The proposed design method not only provides a solid theoretical foundation for the design process of SP network structure cryptography, ensuring systematic and accurate design, but also offers an innovative path for the formal design of cryptographic algorithms.]]></description>
<pubDate>2026/6/4 19:42:18</pubDate>
<category><![CDATA[]]></category>
<author><![CDATA[ZHANG Lei, XU Hongke, XIAO Chaoen, WANG Jianxin, ZHENG Yuzheng]]></author>
</item>
<item>
<title><![CDATA[HTTPFuzzer: Reinforcement Learning Guided Greybox Fuzzing for Web Server Programs]]></title>
<link><![CDATA[http://jcs.iie.ac.cn/xxaqxben/ch/reader/view_abstract.aspx?file_no=20260209&flag=1]]></link>
<description><![CDATA[In the realm of Internet architecture, Web servers function as pivotal elements, facilitating extensive data interchange and intricate business logic execution. The security posture of these servers is intimately tied to the overall stability and resilience of the information systems. Web server programs, as critical components, have received extensive attention from attackers. Consequently, the significance of diligently exploring and promptly remediating vulnerabilities within Web server programs cannot be overstated. The mutation-based greybox fuzzing method is widely used in vulnerability mining of Web server programs. This kind of method usually takes real HTTP messages as "seeds" and mutates the "seeds" to generate test cases. The quality of the test cases depends on the selection of the mutation locations and the scheduling of the mutation operators. Existing methods mainly follow the preset rules in the selection of mutation location and the scheduling of mutation operators. Blindly following the preset rules renders the mutation less targeted and less efficient, resulting in many invalid test cases, which affects the efficiency of fuzzing. To address the aforementioned problems, HTTPFuzzer, a reinforcement learning guided greybox fuzzing method for Web server programs is proposed. HTTPFuzzer’s mutation process is partitioned into two stages. The first stage is the mutation location exploration, where the seeds are segmented first, and the exploration of each segment is transformed into a multi-armed bandit machine problem. During the segment exploration process, according to the code coverage and the responses of the test target, the segments that can make the test target respond correctly and improve the code coverage is selected as the mutation-regions. The second stage is the mutation operator scheduling stage based on Q-learning. This stage aims to improve the code coverage, dynamically adjusts the mutation operator scheduling strategy, and the fuzzer selects better mutation operators according to different mutation-regions. Experimental results demonstrate that HTTPFuzzer can produce high-quality test cases, surpassing benchmark methods by covering more execution paths within a shorter time frame. Furthermore, it is capable of triggering crashes and uncovering potential vulnerabilities in a shorter time.]]></description>
<pubDate>2026/6/4 19:42:18</pubDate>
<category><![CDATA[]]></category>
<author><![CDATA[CHEN Qian, HONG Zheng, JIANG Chuan, ZHANG Guomin, QIN Sujuan, GU Jinbang, CUI Shuai]]></author>
</item>
<item>
<title><![CDATA[Improved PBFT Consensus Algorithm Based on Node Influence and Weighted Aggregation Signature]]></title>
<link><![CDATA[http://jcs.iie.ac.cn/xxaqxben/ch/reader/view_abstract.aspx?file_no=20260210&flag=1]]></link>
<description><![CDATA[The consensus algorithm is the key technology in blockchain to ensure that data is agreed upon. Practical Byzantine Fault Tolerance (PBFT) consensus algorithm can effectively solve the Byzantine Generals problem, and is widely used in distributed systems, blockchain and other scenarios due to its excellent fault tolerance and high efficiency. However, the PBFT consensus algorithm has problems such as lack of reward and punishment mechanism, predictable master node selection and high communication overhead, etc. Aiming at the above-mentioned issues, we proposed an improved PBFT consensus algorithm based on the influence of nodes and weighted aggregation signature. Firstly, a reputation model is designed to dynamically select consensus nodes, with different levels of rewards and penalties based on node types, and a reputation recovery mechanism is designed to prevent “oligopoly” nodes from being generated. Secondly, established the influence assessment mechanism, proposed a novel K-Shell algorithm combining the global and local structure to assess the influence of consensus nodes and identify the key nodes in the consensus network. Meanwhile, a verifiable random function based on the influence of nodes is designed, which improves the probability of key nodes becoming master nodes while making the selection method unpredictable. Finally, we proposed the weight aggregation signature scheme to optimize the consensus process, which reduces the communication overhead and signature volume of the consensus process and improves the consensus efficiency of the algorithm by assigning weights to the nodes for the aggregation signature. Experimental results show that compared with the original PBFT, the average throughput of this paper’s algorithm is improved by 65.7%, and the average delay is reduced by 38.9%, which effectively improves the consensus efficiency of the system. In addition, compared with the typical improved algorithm of PBFT, this paper's algorithm has an obvious performance advantage, and it can be better applied to large-scale consortium chain scenarios.]]></description>
<pubDate>2026/6/4 19:42:18</pubDate>
<category><![CDATA[]]></category>
<author><![CDATA[LIU Lihui, DENG Xiaohong, LIU Yong, SHI Yiran, ZHANG Li]]></author>
</item>
<item>
<title><![CDATA[Design of Verifiable Layered Shuffling Protocol based on Secret Sharing and Its Application Scheme]]></title>
<link><![CDATA[http://jcs.iie.ac.cn/xxaqxben/ch/reader/view_abstract.aspx?file_no=20260211&flag=1]]></link>
<description><![CDATA[The existing shuffling protocols based on secret sharing have some problems, for instance: the existing shuffling protocols mainly focus on the design of theoretical framework, and lack the specific algorithm for each step of the process; Most of the solutions of the existing shuffling protocols using public key are not efficient when dealing with large data sets. The existing shuffling protocol lacks some applicability and is not very practical in some application fields. In view of these limitations, this paper designs a verifiable layered shuffling protocol based on secret sharing. At the same time, in order to combine the specific application scenario, this paper also designed a privacy protection scheme based on the shuffling protocol. In this protocol, an improved share conversion algorithm is constructed on the basis of inadvertent transfer protocol, and the original data set is shuffled without exposing the original data set. The Benes arrangement network is used to realize the shuffling layer, and then the complex shuffling task is divided into multiple sub-tasks that are easy to implement, which improves the processing efficiency of large-scale data. Furthermore, the idea of verifiability is introduced, which ensures the security of the shuffling protocol by allowing participants to confirm that the shuffling process was correctly performed. The correctness of the proposed protocol is analyzed strictly in this paper. The ideal-reality simulation paradigm was used to evaluate the security of the shuffling protocol. The time cost, security and time complexity of algorithm of the shuffling protocol are compared with other protocols. The results of the protocol show that the verifiable layered shuffling protocol based on secret sharing can meet the security standard under the malicious model. It has certain advantages in efficiency when dealing with large-scale data sets. It improves the applicability of the protocol and further promotes its application in the current environment.]]></description>
<pubDate>2026/6/4 19:42:18</pubDate>
<category><![CDATA[]]></category>
<author><![CDATA[ZHANG Yanshuo, MAN Ziqi, ZHOU Xingyu, YANG Yatao, XIE Rongna]]></author>
</item>
<item>
<title><![CDATA[Safety Classification Fine-tuning: A fine-tuning method to improve the output content safety of LLMs]]></title>
<link><![CDATA[http://jcs.iie.ac.cn/xxaqxben/ch/reader/view_abstract.aspx?file_no=20260212&flag=1]]></link>
<description><![CDATA[Instruction-tuned models have been widely applied across various fields and tasks due to their excellent ability to understand and follow instructions. However, this capability is also prone to malicious exploitation, leading the model to generate harmful content. Current methods for enhancing the safety of the output content of instruction-tuned models still have some shortcomings, such as the safety-tuning that can undermine the model's helpfulness, and lack of robust defense against jailbreak attacks, and using pre-trained content moderation models for content filtering can slow down the model's response speed. In response to these challenges, this paper introduces a novel fine-tuning approach known as Safety Classification Fine-tuning (SCFT). The motivation for SCFT is the observation that instruction-tuned models are vulnerable to misuse due to their inability to assess the safety of "instruction-response" pairs. The embedding vector of the EOS token in the hidden state output of the model's final decoding layer, which is used to end sentences, contains the semantic information of the entire sentence and is very suitable to judge the safety of sentences. However, the fundamental structure of the model determines that it does not have classification capabilities. To address this, we have added a new classification head to the model's output layer. This head is trained to classify sentences as "safe" or "unsafe" based on sentences' semantic information, while the model is simultaneously instruction-tuned for general capabilities. The well-trained classification head acts as an internal "discrimination mechanism," controlling the safety of the model's output. This allows the fine-tuned model to actively judge the safety of "instruction-response" pairs during inference and prevent the output of unsafe content. Further analysis reveals that with the "discrimination mechanism", SCFT can unify the training objectives of the model's utility and safety, achieving a better balance between the two. It also maintains the symmetry of knowledge between the model's general capabilities and safety capabilities, and expanding the safety training data to the pre-training data distribution, enhancing the robustness of the model's safety capabilities. Experimental results demonstrate that SCFT is a resource-efficient, end-to-end safety-tuning method. It significantly reduces the Harmfulness Rate of the fine-tuned model by approximately 91%, lowers the average harmfulness score from over 4 points to 1.36 (on a scale of 5, the higher the score, the more harmful the model), and achieves a 0% harmfulness rate in jailbreak attacks, all without increasing additional computing resources or compromising the model's general capabilities.]]></description>
<pubDate>2026/6/4 19:42:18</pubDate>
<category><![CDATA[]]></category>
<author><![CDATA[YU Miao, SUN Lei, HU Cuiyun, ZANG Weifei, GUO Song, HU Peng]]></author>
</item>
<item>
<title><![CDATA[An Efficient Editing Scheme for Large-Scale Blockchain Networks]]></title>
<link><![CDATA[http://jcs.iie.ac.cn/xxaqxben/ch/reader/view_abstract.aspx?file_no=20260213&flag=1]]></link>
<description><![CDATA[Blockchain's immutability has become the cornerstone of trust in decentralized systems, but this feature has also been exploited to store illicit content such as violent, pornographic, and terrorist information. Chameleon hashing, which enables content modification while preserving the hash value through the use of trapdoor information, is recognized as a key technique for achieving blockchain editability. Existing approaches either rely on centralized chameleon hash schemes, granting editing authority to a single entity and undermining blockchain trust, or adopt distributed chameleon hash schemes that maintain trust but fail to detect malicious node behavior during editing. This results in systems operating on erroneous information, producing incorrect outcomes, and wasting computational resources. Moreover, these schemes either retain edited information on the blockchain, contradicting the principles of blockchain editing, or require all network nodes to participate in secure computations to generate trapdoor information and masking data, leading to high communication complexity <i>O</i>(<i>n</i><sup>3</sup>) that limits scalability in large networks. To address these issues, this paper proposes an efficient blockchain editing scheme designed for large-scale network environments. The proposed scheme replaces traditional hash functions with chameleon hash functions in the Merkle root generation process to achieve seamless editing. By utilizing proxy nodes for information aggregation, the scheme reduces communication complexity from <i>O</i>(<i>n</i><sup>3</sup>) to <i>O</i>(<i>n</i><sup>2</sup>). Homomorphic encryption secures trapdoor and masking data, enabling encrypted aggregation and eliminating node dependency on these values. Additionally, the integration of zero-knowledge proofs and commitment mechanisms supports the identification and exclusion of malicious nodes, ensuring stable operation with accurate information. Experiments conducted on 211 blockchain nodes deployed across four cities demonstrate that the proposed scheme outperforms existing methods in efficiency, with its performance advantages becoming increasingly pronounced as network size grows. Fur thermore, rigorous security proofs under the ideal-real model validate the scheme's strong security properties, showcasing its potential for practical application in large-scale blockchain networks.]]></description>
<pubDate>2026/6/4 19:42:18</pubDate>
<category><![CDATA[]]></category>
<author><![CDATA[GAO Yuanpeng, FENG Zhe, LIU Xuefeng, LEI Jing, PEI Qingqi]]></author>
</item>
<item>
<title><![CDATA[Time-Frequency Characteristics Based Multi-Channel Fusion Leakage Detection]]></title>
<link><![CDATA[http://jcs.iie.ac.cn/xxaqxben/ch/reader/view_abstract.aspx?file_no=20260214&flag=1]]></link>
<description><![CDATA[Various information leakages such as power and electromagnetic are generated during the running of cryptographic devices, and The utilization of information leakage poses a serious threat to the actual security of cryptographic device. Leakage detection is an important technology to assess the risk of leakage of cryptographic device, it is to find the evidence of dependency between leakages and sensitive data through hypothesis testing. Detecting only one specific type of information leakage ignores the inherent correlation between multiple information leakages, so it is difficult to fully characterize the actual security of cryptographic devices. Multi-channel fusion leakage detection is a new direction to overcome this technical defect. This paper proposes time-frequency characteristics based multi-channel fusion leakage detection. In both of specific and non-specific scenarios, time-frequency characteristics based multi-channel fusion leakage detection fully utilize the characteristics of hypothesis testing t-test, Hotelling's T2 test, F-test, and Wilk's Lambda test, and combine these four hypothesis testing methods with the time-domain and frequency-domain characteristics of information leakage to deeply explore information leakages related to sensitive data. This paper analyzes the feasibility and applicable scenarios of time-frequency characteristics based multi-channel fusion leakage detection by examining the relationship between multiple factors such as frequency information leakage density, signal-to-noise ratio, dimension, etc and the number of measures required to detect. The experimental results show that the false positive rate of the new method proposed in this paper is reduced by 99.33%-99.97% compared with the existing detection methods when the number of sampling points is the same. In the case of specific test, compared with the existing detection methods, the number of measures required to detect by the new method in this paper is reduced by 15%-52%. In the case of non-specific test, compared with the existing detection methods, the number of measures required to detect by the new method in this paper is reduced by 29%-64%.]]></description>
<pubDate>2026/6/4 19:42:18</pubDate>
<category><![CDATA[]]></category>
<author><![CDATA[FENG Qi, ZHOU Yongbin, MING Jingdian, ZHANG Qian]]></author>
</item>
<item>
<title><![CDATA[A Survey for User Location Privacy Protection Against Inference Attacks in Online Social Networks]]></title>
<link><![CDATA[http://jcs.iie.ac.cn/xxaqxben/ch/reader/view_abstract.aspx?file_no=20260215&flag=1]]></link>
<description><![CDATA[As a new type of mobile application that connects digital space and physical space by location interaction, online social networks can provide users with real-time and convenient online services. When users use the service, their private location information is submitted to the service facing a serious risk of disclosure, including hijacking attacks over mobile devices, man-in-the-middle attacks through network and inference attacks in server-side. This paper was targeted at the potential risk in server environment. It was based on the main characteristics of online social networks and conducted a review study on the defense techniques of both the specific and the combinational inference attack in online social networks. The paper started from the perspective of attack and defense to clearly present the latest progress of online social network users’ location privacy studies. Firstly, based on the in-depth analysis of service model and data characteristics in online social networks, the mechanism of attack models was compared under the traditional specific attack scenario and new combinatorial attack scenario. Then, the classification of user location privacy protection methods was analyzed against inference attacks. For the defense of the specific inference attacks, it was divided into three parts, including data encryption against decryption attacks, identity jamming against re-identification attacks and location distortion against location inference attacks. For the defense of the combinational inference attack, it contains protection solution against three types of same angle combination inference attack, protection solution against three types of two-angle combination inference attack and protection solution against all-angle combination inference attack. By analyzing and summarizing, this paper summarized the differences and characteristics of different inference attack defense schemes, and comprehensively described the evaluation methods and indicators of defense effect. Finally, the research direction of inference attack and hot privacy protection issues was summarized and prospected, which provides ideas and methods for the research in this field.]]></description>
<pubDate>2026/6/4 19:42:18</pubDate>
<category><![CDATA[]]></category>
<author><![CDATA[MA Zhuo, CAO Jiuxin, WANG Qun, XU Shuai, XIA Lingling]]></author>
</item>
<item>
<title><![CDATA[Chinese Text Recognition in Electromagnetic Emission Reconstructed Images Based on Domain Adaptive]]></title>
<link><![CDATA[http://jcs.iie.ac.cn/xxaqxben/ch/reader/view_abstract.aspx?file_no=20260216&flag=1]]></link>
<description><![CDATA[Electromagnetic emission exists in the process of information transmission and display in computer display system. Using TEMPEST technology (Transient Electrical Pulse Analysis Surveillance Technology), radiated electromagnetic information can be easily intercepted. In video images obtained through electromagnetic leakage, the text in the image often contains very important information, which is also the focus of our attention. Therefore, for images obtained through electromagnetic leakage, the recognition of the text area is a crucial task. However, the signal-to-noise ratio of the emitted video signal received by the receiver is very low, and it makes the restored image difficult for effective text recognition. There are few text recognition methods for Chinese text images with low signal-to-noise ratio. In this paper, We propose a CRNN (Convolutional Recurrent Neural Network) text recognition model based on domain adaptation, which uses the unlabeled text images collected in the electromagnetic emission environment as the target domain data, and uses the normal labeled text images as the source domain data. The model combines the Convolutional Neural Network (CNN) with the Domain Discrimination Module(DDM), and then then the semi supervised learning training method is adopted to make the final feature layer extracted by the convolutional neural network be the common features of the target domain dataset with random noise and the normal source domain dataset. As they are common features of both, the impact of various random noise is minimized, and these robust common features can be maximized for subsequent character classification. which improves the accuracy of text recognition in images emitted from target computer. This model was tested on publicly available datasets RCTW-17 and CASIA-10k in the context of electromagnetic leakage restoration, and the evaluation indicators were Precision and Normalized Average Edit Distance (NAED). Compared with mainstream recognition models, The domain adaptation based CRNN has significantly improved the accuracy and normalized average editing distance of text images restored by electromagnetic leakage.]]></description>
<pubDate>2026/6/4 19:42:18</pubDate>
<category><![CDATA[]]></category>
<author><![CDATA[LV Zhiqiang, YU Chao, LI Haiyang, ZHANG Ning]]></author>
</item>
<item>
<title><![CDATA[Survey on Lightweight Virtualization Technology Security]]></title>
<link><![CDATA[http://jcs.iie.ac.cn/xxaqxben/ch/reader/view_abstract.aspx?file_no=20260217&flag=1]]></link>
<description><![CDATA[With the rapid development of lightweight virtualization technology represented by container technology, its position in the cloud computing is becoming more and more important. Lightweight virtualization technology does not create independent operating systems for virtual instances but uses some kernel features to realize the isolation of CPU, memory, network, and file system, which can achieve the full utilization, reasonable allocation, and effective scheduling of hardware resources more efficiently and flexibly. It has brought new technical architectures and operation and maintenance models such as cloud-native to the cloud computing industry. Meanwhile, due to lightweight virtual instances on the same host machine sharing the kernel of the operating system and images in the public repository lack effective security detection, the security isolation mechanism of lightweight virtualization technology is weaker than traditional virtual machine technology. It also brought about new security risks and introduced new security challenges to cloud computing technology, which have received widespread attention in both academia and industry. But its security problems lack systematic research. To understand the security research progress of lightweight virtualization technology, this paper deeply studies and analyzes the security problems and solutions of lightweight virtualization technology. Firstly, we introduce the architecture and application scenarios of lightweight virtualization technology. And we summarize the attack threats of the lightweight virtual instance layer, host machine layer, and hardware layer by the hierarchical model, and generalize the security vulnerability of the image repository and other auxiliary systems. Then, the principle, implementation scheme, types of network attacks that can be defended against, advantages and disadvantages of the existing security defense methods and mechanisms are introduced and analyzed. Finally, this survey paper discusses the future work and suggested security research directions of lightweight virtualization technology. We believe that it is an effective method to improve the security of lightweight virtualization technology by enhancing virtual isolation, ensuring image security detection, and unifying security evaluation criteria.]]></description>
<pubDate>2026/6/4 19:42:19</pubDate>
<category><![CDATA[]]></category>
<author><![CDATA[KONG Tong, WANG Liming, XU Zhen, MA Duohe]]></author>
</item>
<item>
<title><![CDATA[Jointly Exploiting Temporal and Structural Features for Rumor Detection on Social Media]]></title>
<link><![CDATA[http://jcs.iie.ac.cn/xxaqxben/ch/reader/view_abstract.aspx?file_no=20260218&flag=1]]></link>
<description><![CDATA[With the rapids development of the social network, more and more people obtain or share information on social network platforms. Unfortunately, the convenient environment of social network platforms has also provided a new medium for the spread of rumors. The spread of rumor has become a significant challenge that seriously undermines the credibility of information in social network and posts a threat to building a clear cyberspace environment. Automatic rumor detection is essential for timely prevention of rumor spread and maintaining social stability. The existing deep learning-based rumor detection models have been developed based on content characteristics or propagation characteristics including temporal features and structural features. However, most of these detection models either only model the temporal information in rumor propagation or only focus on the structure features of rumor propagation to identify rumors. This limitation cannot learn a comprehensive eigenvector representation well and hinders the performance of rumor detection. To alleviate the above problem, in this paper, we propose a novel graph-based rumor detection model. It combines the power of graph networks and sequence models to jointly model both structural features and temporal patterns in rumor propagation. Specifically, based on the textual features extracted by embedding layer and propagation information, we utilize a time-aware bidirectional gated recurrent unit to explore temporal features and a graph convolutional network to learn structural features. Then, we combine them to make prediction. By doing so, the model can learn a comprehensive representation of rumor characteristics, enabling it to detect rumors with greater accuracy. In addition, the model can effectively alleviate the time mode distortion caused by pruning. To evaluate the performance of the proposed model, we conduct experiments on three real-world rumor detection benchmark datasets. The experimental results show that the proposed method achieves 4.8% average absolute improvements in terms of the accuracy score across all three datasets. Extensive experiments demonstrate the effectiveness of the proposed model for rumor detection.]]></description>
<pubDate>2026/6/4 19:42:19</pubDate>
<category><![CDATA[]]></category>
<author><![CDATA[WEI Lingwei, HU Dou, BAO Yinan, ZHOU Wei, YANG Jinzhu, HU Songlin]]></author>
</item>
<item>
<title><![CDATA[Advance in user identity linkage across online social networks]]></title>
<link><![CDATA[http://jcs.iie.ac.cn/xxaqxben/ch/reader/view_abstract.aspx?file_no=20260219&flag=1]]></link>
<description><![CDATA[With the rapid development of the Internet, social network platforms (also known as online social networks) have become increasingly popular and diversified. In order to make better use of the services provided by each social network platform, users often join multiple social network platforms. Linking the accounts of the same natural person in different social network platforms is called user identity linkage. Through user identity linkage, we can fully understand the user's interests, and greatly enrich the user portrait, which is used in digital marketing and recommendation system. In this paper, by reviewing the different feature types used in the development of the user identity linkage method, a general formal definition of the user identity linkage problem is proposed, which can be applied to various feature types such as attribute, network, content, behavior and any combination of them. Then, according to the two stages of feature extraction and model construction of user identity linkage, the existing user identity linkage methods are classified and analyzed, and different methods are compared and evaluated in terms of performance, computing cost and robustness. Then, different datasets and evaluation indicators used by existing methods are analyzed, the main methods of obtaining datasets are explained, and the reason why there is no publicly recognized benchmark datasets in the field of user identity linkage is given. Finally, the problems and challenges of user identity linkage are discussed, and the future research trend of user identity linkage is forecasted. By proposing a general definition of user identity linkage problem, comparing and analyzing existing user identity linkage methods, discussing existing problems and looking forward to future research trends, this paper analyzes and presents the current situation and future of user identity linkage problem in a clear and structured way, which helps researchers to form a systematic understanding and grasp of related research in this field, and then make more in-depth research work.]]></description>
<pubDate>2026/6/4 19:42:19</pubDate>
<category><![CDATA[]]></category>
<author><![CDATA[XUE Hui, SUN Bo, SI Chengxiang, ZHANG Wei, FANG Jing]]></author>
</item>
<item>
<title><![CDATA[A RLWE-based Three-party Password Authenticated Key Exchange Scheme]]></title>
<link><![CDATA[http://jcs.iie.ac.cn/xxaqxben/ch/reader/view_abstract.aspx?file_no=20260220&flag=1]]></link>
<description><![CDATA[With breakthroughs in quantum theory research, public key cryptosystems based on classical mathematical problems can be cracked by Shor and Grover algorithms with large quantum computers in polynomial time. It becomes very urgent to design cryptographic algorithms that can resist quantum attacks. Many post-quantum algorithms have been studied on the lattice, because the lattice cryptography has some excellent properties such as strong portability and easy-to-implement characteristics, and it has become a current research hotspot. With the purpose of meeting the security and computational efficiency requirements for key exchange in user-user scenarios, this paper proposes a Three-party Password Authenticated Key Exchange (3PAKE) protocol based on the Ring Learning with Errors (RLWE) problem, which introduces the ${\tilde{D}}_{4}$ lattice as reconciliation mechanism, provides identity authentication between the server and two clients through pre-stored passwords, and enables the participants to establish a common secret session key in an insecure channel. In the Bellare Pointcheval Rogaway (BPR) model, it is proved that the protocol has mutual authentication security, weak perfect forward secrecy, session key security and resilience to password guessing attacks. Compared with other RLWE-based authenticated key exchange protocols, the implicit authenticatied scheme significantly reduces the number of hash calculations, and the error reconciliation mechanism allows higher error tolerance and smaller modulus, which leads to a significant reduction in message size and an increase in efficiency and security. Specifically, after balancing the dimensions, modulus, variance, error rate and selecting appropriate parameters, the error rate is reduced to 2<sup>-61</sup> and the modulus is reduced to 12289, which further decreases the amount of calculation and communication complexity. The protocol is implemented in C++ with NFL (NTT-based Fast Lattice) acceleration algorithm, which accelerates the polynomial multiplication and significantly improves the efficiency of the entire protocol. The results in practice show the protocol achieves at most 17x speedup and provides 255-bit quantum security.]]></description>
<pubDate>2026/6/4 19:42:19</pubDate>
<category><![CDATA[]]></category>
<author><![CDATA[WANG Ziliang, GU Xiaozhuo, REN Peixin]]></author>
</item>
</channel>
</rss>