����Ŀ¼

25 April 2023, Volume 9 Issue 2

Previous Issue    Next Issue

Comprehensive Reviews

Survey on vertical federated learning: algorithm, privacy and security

Jinyin CHEN, Rongchang LI, Guohan HUANG, Tao LIU, Haibin ZHENG, Yao CHENG

2023, 9(2):  1-20.  doi:10.11959/j.issn.2096-109x.2023017

Asbtract ( 760 )   HTML ( 163)   PDF (1439KB) ( 1087 )   Knowledge map   

Figures and Tables | References | Related Articles | Metrics

Federated learning （FL） is a distributed machine learning technology that enables joint construction of machine learning models by transmitting intermediate results （e.g., model parameters, parameter gradients, embedding representation, etc.） applied to data distributed across various institutions.FL reduces the risk of privacy leakage, since raw data is not allowed to leave the institution.According to the difference in data distribution between institutions, FL is usually divided into horizontal federated learning （HFL）, vertical federated learning （VFL）, and federal transfer learning （TFL）.VFL is suitable for scenarios where institutions have the same sample space but different feature spaces and is widely used in fields such as medical diagnosis, financial and security of VFL.Although VFL performs well in real-world applications, it still faces many privacy and security challenges.To the best of our knowledge, no comprehensive survey has been conducted on privacy and security methods.The existing VFL was analyzed from four perspectives: the basic framework, communication mechanism, alignment mechanism, and label processing mechanism.Then the privacy and security risks faced by VFL and the related defense methods were introduced and analyzed.Additionally, the common data sets and indicators suitable for VFL and platform framework were presented.Considering the existing challenges and problems, the future direction and development trend of VFL were outlined, to provide a reference for the theoretical research of building an efficient, robust and safe VFL.

Survey on Ethereum phishing detection technology

Zhao CAI, Tao JING, Shuang REN

2023, 9(2):  21-32.  doi:10.11959/j.issn.2096-109x.2023018

Asbtract ( 439 )   HTML ( 95)   PDF (1953KB) ( 617 )   Knowledge map   

Figures and Tables | References | Related Articles | Metrics

With the widespread application of blockchain technology, phishing scams have become a major threat to blockchain platforms.Due to the irreversibility, anonymity, and tamper-proof nature of blockchain transactions, phishing attacks often have a high degree of deception and concealment, causing significant losses to both users and businesses.Ethereum platform, with its smart contract functionality, has attracted many crypto currency investors.However, this widespread popularity has also attracted an influx of criminals, leading to the rise of cybercrime activities.Among them, phishing scams are one of the main forms of fraud on the Ethereum platform.To tackle this issue, researchers have developed Ethereum network phishing identification technology, achieving significant progress in this field.However, there has been relatively little systematic analysis and summary of these research results.The current state of phishing fraud on the Ethereum network was analyzed.Moreover, a comprehensive summary of existing phishing detection datasets and evaluation metrics were provided.On this basis, methods for detecting phishing on Ethereum were reviewed, including those based on transaction information, graph embedding and graph neural networks.Transaction information-based methods are the most common, analyzing information such as input and output addresses and amounts in transaction data to determine whether a transaction is abnormal.Methods based on graph embedding and graph neural networks place more emphasis on analyzing the entire transaction network, constructing a graph structure to analyze the relationships between nodes, and more accurately identifying phishing attacks.In addition, a comparative analysis of the advantages and disadvantages of various methods was conducted, explaining the applicability and limitations of each method.Finally, the challenges facing Ethereum phishing detection were pointed out, and the future research trends for Ethereum phishing detection were predicted.

Papers

Anonymous trust management scheme of VANET based on attribute signature

Min XIAO, Faying MAO, Yonghong HUANG, Yunfei CAO

2023, 9(2):  33-45.  doi:10.11959/j.issn.2096-109x.2023019

Asbtract ( 133 )   HTML ( 25)   PDF (1468KB) ( 290 )   Knowledge map   

Figures and Tables | References | Related Articles | Metrics

Trust management mechanisms can solve the authenticity assessment problem of messages and establish trust between communication entities in the VANET.However, current trust management schemes lack an effective privacy protection mechanism, leading to privacy concerns for vehicles when binding their identity with their trust level.Based on the anonymity of attribute-based signature, an anonymous trust management scheme was proposed for VANET.The trust level of the vehicle was maintained by the trusted authority.The trust level of vehicles and the current time were taken as the attribute identity set of the vehicle, so as to realize the binding of the anonymous vehicle identity with its trust level, which could provide the privacy protection of vehicles.In addition, the timely update of attribute signature private key could resist false reputation attack and the trust level of the vehicle could be verified anonymously while the attribute signature was verified.Furthermore, the pseudonym was used to hide the real identity of the vehicle, enabling only the trusted authority could parse the real identity of vehicles from the pseudonym to update their trust level, and providing legal and secure tracking of the malicious vehicle identity vehicle.Under the general group model, the attribute-based signature scheme is proven to satisfy the security of unforgerability.The security analysis further shows that the scheme protects the identity and location privacy of vehicles, enables only the trusted authority to legally track malicious vehicles, resists false reputation attacks, ensures communication integrity, and resists replay attacks.Performance analysis confirms that the proposed scheme has better computation and communication efficiency than existing schemes.

Method on intrusion detection for industrial internet based on light gradient boosting machine

Xiangdong HU, Lingling TANG

2023, 9(2):  46-55.  doi:10.11959/j.issn.2096-109x.2023020

Asbtract ( 170 )   HTML ( 29)   PDF (1959KB) ( 201 )   Knowledge map   

Figures and Tables | References | Related Articles | Metrics

Intrusion detection is a critical security protection technology in the industrial internet, and it plays a vital role in ensuring the security of the system.In order to meet the requirements of high accuracy and high real-time intrusion detection in industrial internet, an industrial internet intrusion detection method based on light gradient boosting machine optimization was proposed.To address the problem of low detection accuracy caused by difficult-to-classify samples in industrial internet business data, the original loss function of the light gradient boosting machine as a focal loss function was improved.This function can dynamically adjust the loss value and weight of different types of data samples during the training process, reducing the weight of easy-to-classify samples to improve detection accuracy for difficult-to-classify samples.Then a fruit fly optimization algorithm was used to select the optimal parameter combination of the model for the problem that the light gradient boosting machine has many parameters and has great influence on the detection accuracy, detection time and fitting degree of the model.Finally, the optimal parameter combination of the model was obtained and verified on the gas pipeline dataset provided by Mississippi State University, then the effectiveness of the proposed mode was further verified on the water dataset.The experimental results show that the proposed method achieves higher detection accuracy and lower detection time than the comparison model.The detection accuracy of the proposed method on the gas pipeline dataset is at least 3.14% higher than that of the comparison model.The detection time is 0.35s and 19.53s lower than that of the random forest and support vector machine in the comparison model, and 0.06s and 0.02s higher than that of the decision tree and extreme gradient boosting machine, respectively.The proposed method also achieved good detection results on the water dataset.Therefore, the proposed method can effectively identify attack data samples in industrial internet business data and improve the practicality and efficiency of intrusion detection in the industrial internet.

Trust evaluation model of social internet of things based on implicit social relationship

Hongbin ZHANG, Fan FAN, Dongmei ZHAO, Bin LIU, Yan YIN, Jian LIU

2023, 9(2):  56-69.  doi:10.11959/j.issn.2096-109x.2023021

Asbtract ( 141 )   HTML ( 20)   PDF (1930KB) ( 243 )   Knowledge map   

Figures and Tables | References | Related Articles | Metrics

The social internet of things （SIoT） is a new paradigm that combines social networks and the internet of things to enhance network scalability and navigability.However, with increased openness and mobility of intelligent objects, there are rising security risks related to information security and reliable service delivery, which poses a significant challenge to the SIoT.Trust is a critical factor in providing secure and reliable services.Therefore, a new trust assessment model was proposed from the perspective of social relationships.A trust transfer algorithm was proposed based on the explicit social relationships and potential features between nodes.The algorithm can tap and establish the implicit social relationships among nodes to solve the SIoT sparsity and cold start problem.Additionally, the social relations were fine-grained and divided to build a single-relationship sub-network.Multiple single-relationship sub-networks were fused to build a multiple-relationship composite network.Thus, the impact of multiple dynamically changing social relationships among nodes on trust evaluation can be effectively fused.The comprehensive trust assessment of intelligent objects was conducted based on three indicators: direct trust, indirect recommendation trust, and service satisfaction.The experimental results on real smart city datasets demonstrate the model’s robustness in non-sparse and sparse network scenarios.It can effectively evaluate trusted and untrusted objects in the network and improve the accuracy and convergence of trust evaluation.

Distributed user privacy preserving adjustable personalized QoS prediction model for cloud services

Jianlong XU, Jian LIN, Yusen LI, Zhi XIONG

2023, 9(2):  70-80.  doi:10.11959/j.issn.2096-109x.2023022

Asbtract ( 144 )   HTML ( 12)   PDF (1193KB) ( 267 )   Knowledge map   

Figures and Tables | References | Related Articles | Metrics

Personalized quality of service （QoS） prediction is crucial for developing high-quality cloud service system.However, the traditional collaborative filtering method, based on centralized training, presents challenges in protecting user privacy.In order to effectively protect user privacy while obtaining highly accurate prediction effect, a distributed user privacy adjustable personalized QoS prediction model for cloud services （DUPPA） was proposed.The model adopted a “server-multi-user” architecture, in which the server coordinated multiple users, handled multiple users’ requests for uploading model gradients and downloading global model, and maintained global model parameters.To further protect user privacy, a user privacy adjustment strategy was proposed to balance privacy and prediction accuracy by adjusting the initialization proportion of local model parameters and gradient upload proportion.In the local model initialization stage, the user calculated the difference matrix between the local model and the global model, and selected the global model parameters corresponding to the larger elements in the difference matrix to initialize the local model parameters.In the gradient upload stage, the user can select some important gradients to upload to the server to meet the privacy protection requirements of different application scenarios.To evaluate the privacy degree of DUPPA, a data reconstruction attack method was proposed for the distributed matrix factorization model gradient sharing scheme.The experimental results show that when DUPPA sets the gradient upload proportion to 0.1 and the local model parameter initialization proportion to 0.5, the predicted MAE and RMSE are reduced by 1.27% and 0.91%, respectively, compared with the traditional centralized matrix factorization model.Besides, when DUPPA sets the gradient upload proportion to 0.1, the privacy degree is 5 times higher than when the gradient upload proportion is 1.And when DUPPA sets the local model parameter initialization proportion to 0.5, the privacy degree is 3.44 times higher than when the local model parameter initialization proportion is 1.

Progressive active inference method of protocol state machine

Yan PAN, Wei LIN, Yuefei ZHU

2023, 9(2):  81-93.  doi:10.11959/j.issn.2096-109x.2023023

Asbtract ( 151 )   HTML ( 14)   PDF (1963KB) ( 151 )   Knowledge map   

Figures and Tables | References | Related Articles | Metrics

Protocol state machine active inference is a technique that relies on active automata learning.However, the abstraction of the alphabet and the construction of the mapper present critical challenges.Due to the diversity of messages of the same type, the response types of the same type are different, causing the method of regarding the message types as the alphabet will result in the loss of states or state transitions.To address the issue, message types were refined into subtypes according to the different responses and a progressive active inference method was proposed.The proposed method extracted the state fields from the existing protocol data to construct the initial alphabet and the mapper, and obtained the initial state machine based on active automata learning.It then mutated the existing messages to explore the response sequences, which were inconsistent with the current state machine.The mutated message was regarded as a protocol subtype and added to the alphabet, and a new state machine was inferred progressively based on the new alphabet.In order to reduce the interactions, a pre-response query algorithm was proposed based on prefix matching for the caching mechanism in the active automata learning.The ProLearner tool was utilized to evaluate the proposed method in the context of the SMTP and RSTP protocols.It is verified that the pre-response query method can effectively reduce the number of actual interactions, with an average reduction rate of about 10%.

Software diversification method based on binary rewriting

Benwei HE, Yunfei GUO, Yawen WANG, Qingfeng WANG, Hongchao HU

2023, 9(2):  94-103.  doi:10.11959/j.issn.2096-109x.2023024

Asbtract ( 160 )   HTML ( 19)   PDF (3747KB) ( 310 )   Knowledge map   

Figures and Tables | References | Related Articles | Metrics

Software diversity is an effective defense against code-reuse attacks, but most existing software diversification technologies are based on source code.Obtaining program source code may be difficult, while binary files are challenging to disassemble accurately and distinguish between code pointers and data constants.This makes binary file diversification difficult to generate high levels of randomization entropy, and easily compromised by attackers.To overcome these challenges, a binary file oriented software diversification method was proposed based on static binary rewriting technology, namely instruction offset randomization.This method inserted NOP instructions of varying byte lengths before program instructions with a certain probability, reducing the number of unintended gadgets in the program and randomly offsetting the original instruction address.This disrupts the program’s original memory layout and increases the cost of code-reuse attacks.At the same time, an optimization strategy based on hot code was designed for this method.The execution times of basic blocks in binary files were obtained by dynamic pile insertion, so as to adjust the NOP instruction insertion probability in each basic block.The higher the execution frequency, the fewer NOP instructions were inserted into the basic block, which can ensure lower performance overhead and produce higher randomization entropy.In the experimental part, the SPEC benchmark program was used to test the optimized method from the aspects of performance overhead, gadget survival rate and file size.The results show that a 15% insertion probability achieves the best effect, with an average gadget survival rate of less than 1.49%, increasing attackers’ difficulty in reusing the same gadget attack chain.Furthermore, only a 4.1% operation overhead and 7.7% space overhead are added, maintaining high levels of security.

Selfish mining detection scheme based on the characters of transactions

Heli WANG, Qiao YAN

2023, 9(2):  104-114.  doi:10.11959/j.issn.2096-109x.2023025

Asbtract ( 131 )   HTML ( 21)   PDF (2902KB) ( 246 )   Knowledge map   

Figures and Tables | References | Related Articles | Metrics

Selfish mining is an attack strategy in Proof-of-Work based blockchains, where attackers withhold their mined blocks to intercept the awards of other honest miners, resulting in higher profitability.This attack undermines the incentive compatibility of Proof-of-Work.Although there are various studies from researchers in recent years, there is no effective detection scheme as the vulnerability of blockchain in propagation delay makes it challenging to distinguish the attackers from honest miners.A scheme for selfish mining detection was introduced.In this scheme, a state value was assigned to the new block based on transaction amount and the transaction fee of the block.By analyzing the relationship between state value and transaction characteristics, we can determine if the block was mined by a selfish miner.The scheme is verified by an experiment with an accuracy of 86.02%.

Binary program taint analysis optimization method based on function summary

Pan YANG, Fei KANG, Hui SHU, Yuyao HUANG, Xiaoshao LYU

2023, 9(2):  115-131.  doi:10.11959/j.issn.2096-109x.2023026

Asbtract ( 224 )   HTML ( 26)   PDF (2562KB) ( 269 )   Knowledge map   

Figures and Tables | References | Related Articles | Metrics

Taint analysis is a popular software analysis method, which has been widely used in the field of information security.Most of the existing binary program dynamic taint analysis frameworks use instruction-level instrumentation analysis methods, which usually generate huge performance overhead and reduce the program execution efficiency by several times or even dozens of times.This limits taint analysis technology’s wide usage in complex malicious samples and commercial software analysis.An optimization method of taint analysis based on function summary was proposed, to improve the efficiency of taint analysis, reduce the performance loss caused by instruction-level instrumentation analysis, and make taint analysis to be more widely used in software analysis.The taint analysis method based on function summary used function taint propagation rules instead of instruction taint propagation rules to reduce the number of data stream propagation analysis and effectively improve the efficiency of taint analysis.For function summary, the definition of function summary was proposed.And the summary generation algorithms of different function structures were studied.Inside the function, a path-sensitive analysis method was designed for acyclic structures.For cyclic structures, a finite iteration method was designed.Moreover, the two analysis methods were combined to solve the function summary generation of mixed structure functions.Based on this research, a general taint analysis framework called FSTaint was designed and implemented, consisting of a function summary generation module, a data flow recording module, and a taint analysis module.The efficiency of FSTaint was evaluated in the analysis of real APT malicious samples, where the taint analysis efficiency of FSTaint was found to be 7.75 times that of libdft, and the analysis efficiency was higher.In terms of accuracy, FSTaint has more accurate and complete propagation rules than libdft.

High-performance directional fuzzing scheme based on deep reinforcement learning

Tian XIAO, Zhihao JIANG, Peng TANG, Zheng HUANG, Jie GUO, Weidong QIU

2023, 9(2):  132-142.  doi:10.11959/j.issn.2096-109x.2023027

Asbtract ( 287 )   HTML ( 48)   PDF (3151KB) ( 472 )   Knowledge map   

Figures and Tables | References | Related Articles | Metrics

With the continuous growth and advancement of the Internet and information technology, continuous growth and advancement of the Internet and information technology.Nevertheless, these applications’ vulnerabilities pose a severe threat to information security and users’ privacy.Fuzzing was widely used as one of the main tools for automatic vulnerability detection due to its ease of vulnerability recurrence and low false positive errors.It generates test cases randomly and executes the application by optimization in terms of coverage or sample generation to detect deeper program paths.However, the mutation operation in fuzzing is blind and tends to make the generated test cases execute the same program path.Consequently, traditional fuzzing tests have problems such as low efficiency, high randomness of inputs generation and limited pertinence of the program structure.To address these problems, a directional fuzzing based on deep reinforcement learning was proposed, which used deep reinforcement learning networks with information obtained by staking program to guide the selection of the inputs.Besides, it enabled fast approximation and inspection of the program paths that may exist vulnerabilities.The experimental results showed that the proposed approach had better performance than the popular fuzzing tools such as AFL and AFLGO in terms of vulnerability detection and recurrence on the LAVA-M dataset and real applications like LibPNG and Binutils.Therefore, the approach can provide support for further vulnerability mining and security research.

Dynamic multi-keyword searchable encryption scheme

Chenghao YUAN, Yong LI, Shuang REN

2023, 9(2):  143-153.  doi:10.11959/j.issn.2096-109x.2023028

Asbtract ( 212 )   HTML ( 30)   PDF (1240KB) ( 172 )   Knowledge map   

Figures and Tables | References | Related Articles | Metrics

Users of cloud storage often outsource their data to cloud servers to save local storage resources.However, cloud storage data is out of the user’s physical control, which may lead to malicious theft or disclosure of private data on cloud.At present, dynamic searchable encryption schemes are mostly based on symmetric searchable encryption, which need to establish a secure key sharing channel in advance, and cannot be directly applied to data sharing in cloud storage scenarios.For the secure sharing scenario of multi-party data in cloud storage, a dynamic multi-keyword searchable encryption scheme was proposed.The forward index was constructed by a cuckoo filter to enable data owners to perform documents and index dynamic updates.The conjunctive multi-keywords search was supported by combining bilinear pairing with Lagrangian interpolation polynomials.To reduce the computational overhead in the ciphertext retrieval phase, a combination of inverted index and forward index was constructed to improve the retrieval efficiency of cloud server.This scheme is provably secure with the indistinguishability in adaptively chosen keyword attack under decision linear Diffie-Hellman problem.Simulation experiments were conducted to analyze the execution efficiency of the scheme for keyword search and index update in different datasets.The results show that the scheme effectively avoids the linear correlation between the retrieval time and the number of ciphertexts, and reduces the computational overhead in the update operation with a large amount of data.

Security evaluation for parameters of SIMON-like cipher based on neural network distinguisher

Zezhou HOU, Jiongjiong REN, Shaozhen CHEN

2023, 9(2):  154-163.  doi:10.11959/j.issn.2096-109x.2023029

Asbtract ( 91 )   HTML ( 10)   PDF (1627KB) ( 123 )   Knowledge map   

Figures and Tables | References | Related Articles | Metrics

The neural distinguisher is a new tool widely used in crypto analysis of some ciphers.For SIMON-like block ciphers, there are multiple choices for their parameters, but the reasons for designer’s selection remain unexplained.Using neural distinguishers, the security of the parameters （a,b,c） of the SIMON-like with a block size of 32 bits was researched, and good choices of parameters were given.Firstly, using the idea of affine equivalence class proposed by K?lbl et al.in CRYPTO2015, these parameters can be divided into 509 classes.And 240 classes which satisfied gcd（a-b,2）=1 were mainly researched.Then a SAT/SMT model was built to help searching differential characteristics for each equivalent class.From these models, the optimal differential characteristics of SIMON-like was obtained.Using these input differences of optimal differential characteristics, the neural distinguishers were trained for the representative of each equivalence class, and the accuracy of the distinguishers was saved.It was found that 20 optimal parameters given by K?lbl et al.cannot make the neural distinguishers the lowest accuracy.On the contrary, there were 4 parameters, whose accuracy exceeds 80%.Furthermore, the 4 parameters were bad while facing neural distinguishers.Finally, comprehensively considering the choice of K?lbl et al.and the accuracy of different neural distinguishers, three good parameters, namely （6,11,1）,（1,8,3）, and（6,7,5） were given.

Research on construction technology of artificial intelligence security knowledge graph

Xiaochen SHEN, Yinhui GE, Bo CHEN, Ling YU

2023, 9(2):  164-174.  doi:10.11959/j.issn.2096-109x.2023030

Asbtract ( 316 )   HTML ( 55)   PDF (2879KB) ( 544 )   Knowledge map   

Figures and Tables | References | Related Articles | Metrics

As a major strategic technology, artificial intelligence is developing rapidly while bringing numerous security risks.Currently, security data for artificial intelligence is collected from disparate sources and lacks standardized description, making it difficult to integrate and analyze effectively.To address this issue, a method for constructing an artificial intelligence security knowledge graph was proposed.The knowledge graph was used to integrate the current multi-source heterogeneous data, scientifically represent complex relationships of the data, mine potential value and form a domain knowledge base.In view of the diversity and correlation of concepts in the field of artificial intelligence security, a hierarchical structure of artificial intelligence security ontology was proposed to make the ontology structure more diversified and extensible, provide rule constraints for the process of knowledge graph construction, and form an artificial intelligence security knowledge base.To effectively utilize feature information and reduce noise interference, named entity recognition algorithm based on BiLSTM-CRF and relationship extraction algorithm based on CNN-ATT were adopted for information extraction.The constructed artificial intelligence security dataset was then used to verify the performance of the algorithm.Based on the proposed ontology, the multi-level visualization results of the artificial intelligence security knowledge graph were presented in 3D effect, effectively connecting the multi-source security data information.The experimental results show that the constructed knowledge graph meets the multi-dimensional evaluation criteria of accuracy, consistency, completeness, and timeliness, providing knowledge support for artificial intelligence security research.Overall, the proposed method can help address the complexity and heterogeneity of security data in artificial intelligence and provide a more standardized, integrated approach to knowledge representation and analysis.

Traditional guidance mechanism based deep robust watermarking

Xuejing GUO, Yixiang FANG, Yi ZHAO, Tianzhu ZHANG, Wenchao ZENG, Junxiang WANG

2023, 9(2):  175-183.  doi:10.11959/j.issn.2096-109x.2023031

Asbtract ( 141 )   HTML ( 21)   PDF (10353KB) ( 88 )   Knowledge map   

Figures and Tables | References | Related Articles | Metrics

With the development of network and multimedia technology, multimedia data has gradually become a key source of information for people, making digital media the primary battlefield for copyright protection and anti-counterfeit traceability.Digital watermarking techniques have been widely studied and recognized as important tools for copyright protection.However, the robustness of conventional digital watermarking methods is limited as sensitive digital media can easily be affected by noise and external interference during transmission.Then the existing powerful digital watermarking technology’s comprehensive resistance to all forms of attacks must be enhanced.Moreover, the conventional robust digital watermarking algorithm’s generalizability across a variety of image types is limited due to its embedding method.Deep learning has been widely used in the development of robust digital watermarking systems due to its self-learning abilities.However, current initialization techniques based on deep neural networks rely on random parameters and features, resulting in low-quality model generation, lengthy training times, and potential convergence issues.To address these challenges, a deep robust digital watermarking algorithm based on a traditional bootstrapping mechanism was proposed.It combined the benefits of both traditional digital watermarking techniques and deep neural networks, taking into account their learning abilities and robust characteristics.The algorithm used the classic robust digital watermarking algorithm to make watermarked photos, and the constructed feature guaranteed the resilience of traditional watermarked images.The final dense image was produced by fusing the conventionally watermarked image with the deep network using the U-Net structure.The testing results demonstrate that the technique can increase the stego image’s resistance to various attacks and provide superior visual quality compared to the conventional algorithm.

Application and risk response of deep synthesis technology

Jingwen LI, Yawen LI

2023, 9(2):  184-190.  doi:10.11959/j.issn.2096-109x.2023032

Asbtract ( 286 )   HTML ( 95)   PDF (823KB) ( 278 )   Knowledge map   

Figures and Tables | References | Related Articles | Metrics

Deep synthesis technology is a widely used application in the field of artificial intelligence, particularly in intelligent services, audio and video production, media communication and information services.A basic research was conducted on the principle and implementation of deep synthesis technology.And the application of deep synthesis technology were analyzed, such as robot writing, voice synthesis, intelligent customer service, voice replication, face synthesis, posture manipulation, virtual characters and virtual scenes.Based on the analysis of the characteristics and development trend of the application of deep synthesis, the risks brought by the application of deep synthesis technology were summarized, such as the generation and dissemination of false information, infringement of the legitimate rights and interests of others, utilization by other illegal and criminal activities, and impact on national security.The measures taken by the EU, United States and China to govern the application of deep synthesis technology were studied and the countermeasures were recommended such as building a comprehensive regulatory rule system, strengthening technical supervision, and improving risk prevention awareness across society.Furthermore, the need for intelligence, authenticity, and universality in the application of deep synthesis technology was highlighted.Based on the analysis of the characteristics and development trends of deep synthesis technology, a comprehensive insight into the potential risks and countermeasures associated with its application governance were provided.