This article elucidates the research progress in enhancing container security through the isolation of system calls. The article firstly outlines the development background of containerization technology and its major security challenges. Subsequently, an indepth analysis is conducted on the role of system call isolation in enhancing the security of containers, including the techniques of limiting the system calls of containerized applications to reduce the attack surface, and leveraging operating system middleware and hardware protection mechanisms to accomplish the isolation and protection of containers. By comparing the implementation principles, performance, and their effects on isolation, reduction of attack surfaces, and data protection, the article reveals the advantages and limitations of system call isolation technologies in enhancing container security.

Large language model (LLM), with its powerful feature learning ability, the ability to recognize complex patterns, and generalization ability, has paved the way for innovative and powerful methods in network attack detection. Firstly, this paper elaborates on the technical advantages of LLM in network attack detection and proposes a corresponding technical framework. Then, drawing on existing literature, the application status of LLM in network attack detection is reviewed from three aspects: processing original security data, extracting threat features, correlation analysis, and identifying threats in the target environment. Furthermore, the problems and challenges associated with network threat detection using LLM are analyzed. Lastly, the paper outlines the future research directions for network attack detection technology leveraging LLM. This paper aims to provide references for the further development of network attack detection technology based on LLM in the field of network security.

Aiming at the problems of inefficient exploration, nonguided test case generation, and poor constraintsolving tenacity in current hybrid fuzzy testing frameworks for smart contracts, this paper proposes an improved hybrid fuzzy detection framework DEST.The model integrates the advantages of fuzzy testing and symbolic execution methods to efficiently detect smart contracts, incorporates the differential evolution (DE) algorithm to optimize the quality of test cases and global search capability, and learns SMT script features through LSTM framework to predict the solving time. The DEST model uses the differential evolutionary (DE) algorithm to optimize the quality of test cases and global search capability, and learns SMT script features through LSTM framework to predict the solving time,thereby improving the solving efficiency of symbolic execution. Experiments show that the DEST model improves vulnerability detection by 9.42% and average code coverage by 3.6% over the stateoftheart benchmark model.

To enable scalable IoT systems, edge computing, as a new decentralized model, is introduced into IoT scenarios. Zero trust architecture (ZTA) is wellsuited for cloudedgeend systems with blurred boundaries, offering continuous dynamic authentication and improved security. Due to their lightweight and unclonable properties, physical unclonable functions (PUFs) are often used to generate hardware fingerprint identities for devices. PUFs exploit inherent randomness introduced during hardware fabrication processes to generate unique and nonpredictable challengeresponse pairs. If an attacker collects many plaintext CRPs during continuous authentication, he may model and predict future responses, enabling machine learning attacks. This paper proposes a PUFbased authentication solution (PAMLCA). It enhances privacy protection against machine learning attacks by leveraging oblivious pseudorandom function techniques to obfuscate CRP transmission. The solution combines static and continuous multilayer dynamic verification protocols, limiting implicit trust domains within a session. Security analysis and performance comparisons demonstrate that PAMLCA offers better security, functionality, communication, and computational efficiency compared to other related solutions.

The power system features numerous nodes and heavy traffic. Current quantum key distribution have insufficient key generation rates to meet the encryption requirements of power system services. This paper proposes a dynamic feedbackbased quantum key adjustment scheme to address the security challenges in power system under limited quantum key resources in crossdomain key pools. The scheme consists of two phases, corresponding to the dynamic adjustment of key pool input and output. The feedback mechanism is applied to maintain equilibrium. During the dynamic adjustment of input, realtime selection and optimization of quantum key distribution protocols are carried out based on environmental factors to improve the realtime input rate of quantum keys. During the dynamic adjustment of output, the allocation and utilization of quantum key resources are settled to maximize the overall security level of data within the sampling time. Feedback on data security is reported to the input phase to ensure that the encryption can reach the predetermined lower limit of overall security level. Experimental results show that the proposed scheme achieves an average gain of 12.59% in overall service security under highsecurityrisk conditions, outperforming conventional methods.

In the current era of digital economy, where globalization and informatization are deeply integrated, blockchain technology due to its inherent features of decentralization, immutability, and transparency has offered innovative solutions for secure data storage, value transfer, and trust building. Given the system complexity brought about by the distributed and decentralized nature of blockchain systems, conducting effectiveness tests is of particular importance. Taking this as a starting point, this research deeply analyzes blockchain technology and existing testing tools at home and abroad, and constructs a universal testing platform for heterogeneous and multimodal blockchain systems, with a focus on technical dimensions such as compatibility, universality, scalability, stability, reliability, and security. The aim is to promote research, facilitate product development, and enable realworld applications through testing, thereby accelerating the pace of innovation in blockchain technology.

With the continuous advancement of artificial intelligence and big data technologies, the threshold for making fake videos has been significantly reduced. Therefore, identifying whether a video has been tampered with is crucial for ensuring the authenticity and credibility of the information. Current mainstream video forgery detection methods rely on convolutional neural networks, which exhibit limited capability in capturing temporal dependencies and lack comprehensive understanding of global temporal patterns. To address this issue, this paper proposes a learnable Ptuning based method for video object removal tamper detection and localization. Firstly, the prior knowledge of the pretrained model is fully mined by learnable Ptuning, and multiview features such as spatial, temporal and highfrequency are efficiently extracted. Secondly, a multiscale feature interaction module is proposed to accurately capture the tampering traces from finegrained to coarsegrained through multiscale convolution operation and twostep decomposition strategy. Furthermore, a multiview fusion attention module is designed to significantly enhance the information sharing and fusion ability among multiview features via the crossview interaction mechanism. Experimental results demonstrate that the proposed method outperforms existing detection methods in both the time domain and the spatial domain.

The reliance of generative artificial intelligence training on massive volumes of copyrighted works has given rise to increasingly significant risks of copyright infringement. Jurisdictions such as the European Union, the United States, and Japan have introduced regulatory responses, including innovative rules on text and data mining exceptions. Although allowing the use of copyrighted works for data training has become a general theoretical consensus in China, there remains considerable controversy over the specific pathways to compliance. This article argues for the establishment of a copyright open licensing mechanism for data training, replacing individualized authorization with voluntary public declarations, and incentivizing right holders’ participation through fair benefit allocation and transparent regulatory safeguards. This approach aims to strike a dynamic balance between technological innovation and copyright protection. Given the automatic protection and vast quantity of copyrighted works, the legal effect of publicity of open licensing declarations should be expressly recognized to protect bona fide thirdparty reliance. Additionally, right holders should be permitted to grant collective licenses for series or sets of works to better accommodate the dataintensive utilization demands in the era of artificial intelligence.

The use of facial recognition by administrative organs has special characteristics in terms of the source of usage permissions, usage purposes, usage methods, etc., which puts forward a practical need for differentiated norms for the use of facial recognition by administrative organs. However, at present, there is no special regulation on the personal information processing behavior of administrative organs in China, and the relevant legal provisions are relatively rough and vague. Therefore, in response to the existing problems and based on the requirements of building a rule of law government, it is proposed to choose an appropriate legal regulatory model, restrict facial recognition application subjects through prior approval, grant facial recognition application permissions in stages, facilitate participation in supervision channels and improve information transparency, and carry out differentiated accountability for different types of administrative actions, so as to achieve the legalization of the institutions, functions, authorities, procedures, and responsibilities of administrative agencies using facial recognition.

In order to cope with the potential risks and regulatory challenges brought by the rapid development of artificial intelligence technology, exploring how to stimulate technological innovation while ensuring public safety has become the core topic in the current reform of regulatory systems. This paper proposes an institutional analysis method to systematically analyze the balancing role of the EU’s artificial regulatory sandbox system in promoting technological innovation and ensuring public safety. The research findings reveal that although the system has the positive significance of reducing the risk cost of enterprises and improving the effectiveness of supervision, it also has the problems of limited application scenarios, limited scope of liability exemption and the emergence of power rentseeking in the access process. Based on the experience of the EU, China can advance the exploration of the artificial intelligence regulatory sandbox system in the following aspects: coordinating technological innovation and public safety, and establishing a centrallocal coordinated regulatory agency; optimizing the application, evaluation, testing, reporting and exit process; improving consumers’ risk tolerance, and safeguarding their rights to know and to be compensated; and through the establishment of regulatory chat rooms and institutional incentives to promote the effective implementation of the system.

With the continuous evolution of hacking technology, the iterative upgrades of malicious code variants have been acclerating and the number of malicious codes has exploded. How to rapidly and accurately detect malicious code has become a challenging research hotspot in the realm of cybersecurity. The unique gating mechanism of long shortterm memory network (LSTM) can selectively retain important historical information. Moreover, it demonstrates excellent performance for the sequential dependence of data on time series, which can effectively solve the problem of gradient vanishing or gradient explosion that may occur when traditional RNNs deal with such problems. This distinctive sequential processing capability of LSTM is particularly important for malware detection, thus learning to its extensive application in this area. This paper comprehensively sorts out and summarizes the application of LSTM in malicious code detection from five aspects, including the detection method of malicious code, the basic model and variants of LSTM, the application of LSTM in malicious code detection, the performance analysis of LSTM in malicious code detection, and the future development direction of LSTM in the field of malicious code detection, aiming to facilitating further research and improvement of existing methods for malicious code detection.