Zenghui Yuan
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Previous Organizations / Institutes
The organizations / institutes from the scholar's other articles in the past ten years
Research Subjects
Zenghui Yuan has a total of 20 co-authors. Here are the top co-authors by number of shared publications:
- Pan Zhou
Huazhong University of Science and Technology
7 shared publications - Lichao Sun
Lehigh University, Bethlehem, USA
4 shared publications - Jiawen Shi
Center on Big Data Security School of Cyber Science and Engineering Hubei Key Laboratory of Distributed System Security Hubei Engineering Research Huazhong University of Science and Technology
2 shared publications - Neil Zhenqiang Gong
Duke University, Durham, NC, USA
2 shared publications - Yangming Xu
Center on Big Data Security School of Cyber Science and Engineering Hubei Key Laboratory of Distributed System Security Hubei Engineering Research Huazhong University of Science and Technology
2 shared publications - Cheng Wang
Huazhong University of Science and Technology, Wuhan, China
1 shared publication
Latest Publications
- Open Access
An Image Is Worth Ten Thousand Words: Verbose-Text Induction Attacks on VLMs
With the remarkable success of Vision-Language Models (VLMs) on multimodal tasks, concerns regarding their deployment efficiency have become increasingly prominent. In particular, the number of tokens consumed during the generation process has emerged as a key evaluation metric.Prior studies have shown that specific inputs can induce VLMs to generate lengthy outputs with low information density, which significantly increases energy consumption, latency, and token costs. However, existing methods simply delay the occurrence of the EOS token to implicitly prolong output, and fail to directly maximize the output token length as an explicit optimization objective, lacking stability and controllability.To address these limitations, this paper proposes a novel verbose-text induction attack (VTIA) to inject imperceptible adversarial perturbations into benign images via a two-stage framework, which identifies the most malicious prompt embeddings for optimizing and maximizing the output token of the perturbed images.Specifically, we first perform adversarial prompt search, employing reinforcement learning strategies to automatically identify adversarial prompts capable of inducing the LLM component within VLMs to produce verbose outputs. We then conduct vision-aligned perturbation optimization to craft adversarial examples on input images, maximizing the similarity between the perturbed image's visual embeddings and those of the adversarial prompt, thereby constructing malicious images that trigger verbose text generation. Comprehensive experiments on four popular VLMs demonstrate that our method achieves significant advantages in terms of effectiveness, efficiency, and generalization capability.
Open AccessAn Image Is Worth Ten Thousand Words: Verbose-Text Induction Attacks on VLMs
With the remarkable success of Vision-Language Models (VLMs) on multimodal tasks, concerns regarding their deployment efficiency have become increasingly prominent. In particular, the number of tokens consumed during the generation process has emerged as a key evaluation metric.Prior studies have shown that specific inputs can induce VLMs to generate lengthy outputs with low information density, which significantly increases energy consumption, latency, and token costs. However, existing methods simply delay the occurrence of the EOS token to implicitly prolong output, and fail to directly maximize the output token length as an explicit optimization objective, lacking stability and controllability.To address these limitations, this paper proposes a novel verbose-text induction attack (VTIA) to inject imperceptible adversarial perturbations into benign images via a two-stage framework, which identifies the most malicious prompt embeddings for optimizing and maximizing the output token of the perturbed images.Specifically, we first perform adversarial prompt search, employing reinforcement learning strategies to automatically identify adversarial prompts capable of inducing the LLM component within VLMs to produce verbose outputs. We then conduct vision-aligned perturbation optimization to craft adversarial examples on input images, maximizing the similarity between the perturbed image's visual embeddings and those of the adversarial prompt, thereby constructing malicious images that trigger verbose text generation. Comprehensive experiments on four popular VLMs demonstrate that our method achieves significant advantages in terms of effectiveness, efficiency, and generalization capability.
- Open Access
Merge Hijacking: Backdoor Attacks to Model Merging of Large Language Models
Model merging for Large Language Models (LLMs) directly fuses the parameters of different models finetuned on various tasks, creating a unified model for multi-domain tasks. However, due to potential vulnerabilities in models available on open-source platforms, model merging is susceptible to backdoor attacks. In this paper, we propose Merge Hijacking, the first backdoor attack targeting model merging in LLMs. The attacker constructs a malicious upload model and releases it. Once a victim user merges it with any other models, the resulting merged model inherits the backdoor while maintaining utility across tasks. Merge Hijacking defines two main objectives-effectiveness and utility-and achieves them through four steps. Extensive experiments demonstrate the effectiveness of our attack across different models, merging algorithms, and tasks. Additionally, we show that the attack remains effective even when merging real-world models. Moreover, our attack demonstrates robustness against two inference-time defenses (Paraphrasing and CLEANGEN) and one training-time defense (Fine-pruning).
Open AccessMerge Hijacking: Backdoor Attacks to Model Merging of Large Language Models
Model merging for Large Language Models (LLMs) directly fuses the parameters of different models finetuned on various tasks, creating a unified model for multi-domain tasks. However, due to potential vulnerabilities in models available on open-source platforms, model merging is susceptible to backdoor attacks. In this paper, we propose Merge Hijacking, the first backdoor attack targeting model merging in LLMs. The attacker constructs a malicious upload model and releases it. Once a victim user merges it with any other models, the resulting merged model inherits the backdoor while maintaining utility across tasks. Merge Hijacking defines two main objectives-effectiveness and utility-and achieves them through four steps. Extensive experiments demonstrate the effectiveness of our attack across different models, merging algorithms, and tasks. Additionally, we show that the attack remains effective even when merging real-world models. Moreover, our attack demonstrates robustness against two inference-time defenses (Paraphrasing and CLEANGEN) and one training-time defense (Fine-pruning).
- Open Access
BadToken: Token-level Backdoor Attacks to Multi-modal Large Language Models
Multi-modal large language models (MLLMs) extend large language models (LLMs) to process multi-modal information, enabling them to generate responses to image-text inputs. MLLMs have been incorporated into diverse multi-modal applications, such as autonomous driving and medical diagnosis, via plug-and-play without fine-tuning. This deployment paradigm increases the vulnerability of MLLMs to backdoor attacks. However, existing backdoor attacks against MLLMs achieve limited effectiveness and stealthiness. In this work, we propose BadToken, the first token-level backdoor attack to MLLMs. BadToken introduces two novel backdoor behaviors: Token-substitution and Token-addition, which enable flexible and stealthy attacks by making token-level modifications to the original output for backdoored inputs. We formulate a general optimization problem that considers the two backdoor behaviors to maximize the attack effectiveness. We evaluate BadToken on two open-source MLLMs and various tasks. Our results show that our attack maintains the model's utility while achieving high attack success rates and stealthiness. We also show the real-world threats of BadToken in two scenarios, i.e., autonomous driving and medical diagnosis. Furthermore, we consider defenses including fine-tuning and input purification. Our results highlight the threat of our attack.
Open AccessBadToken: Token-level Backdoor Attacks to Multi-modal Large Language Models
Multi-modal large language models (MLLMs) extend large language models (LLMs) to process multi-modal information, enabling them to generate responses to image-text inputs. MLLMs have been incorporated into diverse multi-modal applications, such as autonomous driving and medical diagnosis, via plug-and-play without fine-tuning. This deployment paradigm increases the vulnerability of MLLMs to backdoor attacks. However, existing backdoor attacks against MLLMs achieve limited effectiveness and stealthiness. In this work, we propose BadToken, the first token-level backdoor attack to MLLMs. BadToken introduces two novel backdoor behaviors: Token-substitution and Token-addition, which enable flexible and stealthy attacks by making token-level modifications to the original output for backdoored inputs. We formulate a general optimization problem that considers the two backdoor behaviors to maximize the attack effectiveness. We evaluate BadToken on two open-source MLLMs and various tasks. Our results show that our attack maintains the model's utility while achieving high attack success rates and stealthiness. We also show the real-world threats of BadToken in two scenarios, i.e., autonomous driving and medical diagnosis. Furthermore, we consider defenses including fine-tuning and input purification. Our results highlight the threat of our attack.
- Open Access
- 1 January 2025
- Published by Association for Computational Linguistics (ACL)
- p. 32688-32703
Merge Hijacking: Backdoor Attacks to Model Merging of Large Language Models
Zenghui Yuan, Yangming Xu, Jiawen Shi, Pan Zhou, Lichao Sun. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.
- 1 January 2025
- Published by Association for Computational Linguistics (ACL)
- p. 32688-32703
Open AccessMerge Hijacking: Backdoor Attacks to Model Merging of Large Language Models
Zenghui Yuan, Yangming Xu, Jiawen Shi, Pan Zhou, Lichao Sun. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.
- Top Cited PapersOpen Access
- 2 December 2024
- Published by Association for Computing Machinery (ACM)
- p. 660-674
Optimization-based Prompt Injection Attack to LLM-as-a-Judge
LLM-as-a-Judge uses a large language model (LLM) to select the best response from a set of candidates for a given question. LLM-as-a-Judge has many applications such as LLM-powered search, reinforcement learning with AI feedback (RLAIF), and tool selection. In this work, we propose JudgeDeceiver, an optimization-based prompt injection attack to LLM-as-a-Judge. JudgeDeceiver injects a carefully crafted sequence into an attacker-controlled candidate response such that LLM-as-a-Judge selects the candidate response for an attacker-chosen question no matter what other candidate responses are. Specifically, we formulate finding such sequence as an optimization problem and propose a gradient based method to approximately solve it. Our extensive evaluation shows that JudgeDeceive is highly effective, and is much more effective than existing prompt injection attacks that manually craft the injected sequences and jailbreak attacks when extended to our problem. We also show the effectiveness of JudgeDeceiver in three case studies, i.e., LLM-powered search, RLAIF, and tool selection. Moreover, we consider defenses including known-answer detection, perplexity detection, and perplexity windowed detection. Our results show these defenses are insufficient, highlighting the urgent need for developing new defense strategies.
- 2 December 2024
- Published by Association for Computing Machinery (ACM)
- p. 660-674
Top Cited PapersOpen AccessOptimization-based Prompt Injection Attack to LLM-as-a-Judge
LLM-as-a-Judge uses a large language model (LLM) to select the best response from a set of candidates for a given question. LLM-as-a-Judge has many applications such as LLM-powered search, reinforcement learning with AI feedback (RLAIF), and tool selection. In this work, we propose JudgeDeceiver, an optimization-based prompt injection attack to LLM-as-a-Judge. JudgeDeceiver injects a carefully crafted sequence into an attacker-controlled candidate response such that LLM-as-a-Judge selects the candidate response for an attacker-chosen question no matter what other candidate responses are. Specifically, we formulate finding such sequence as an optimization problem and propose a gradient based method to approximately solve it. Our extensive evaluation shows that JudgeDeceive is highly effective, and is much more effective than existing prompt injection attacks that manually craft the injected sequences and jailbreak attacks when extended to our problem. We also show the effectiveness of JudgeDeceiver in three case studies, i.e., LLM-powered search, RLAIF, and tool selection. Moreover, we consider defenses including known-answer detection, perplexity detection, and perplexity windowed detection. Our results show these defenses are insufficient, highlighting the urgent need for developing new defense strategies.
- Open Access
The Security and Privacy of Mobile Edge Computing: An Artificial Intelligence Perspective
Mobile Edge Computing (MEC) is a new computing paradigm that enables cloud computing and information technology (IT) services to be delivered at the network's edge. By shifting the load of cloud computing to individual local servers, MEC helps meet the requirements of ultralow latency, localized data processing, and extends the potential of Internet of Things (IoT) for end-users. However, the crosscutting nature of MEC and the multidisciplinary components necessary for its deployment have presented additional security and privacy concerns. Fortunately, Artificial Intelligence (AI) algorithms can cope with excessively unpredictable and complex data, which offers a distinct advantage in dealing with sophisticated and developing adversaries in the security industry. Hence, in this paper we comprehensively provide a survey of security and privacy in MEC from the perspective of AI. On the one hand, we use European Telecommunications Standards Institute (ETSI) MEC reference architecture as our based framework while merging the Software Defined Network (SDN) and Network Function Virtualization (NFV) to better illustrate a serviceable platform of MEC. On the other hand, we focus on new security and privacy issues, as well as potential solutions from the viewpoints of AI. Finally, we comprehensively discuss the opportunities and challenges associated with applying AI to MEC security and privacy as possible future research directions.
Open AccessThe Security and Privacy of Mobile Edge Computing: An Artificial Intelligence Perspective
Mobile Edge Computing (MEC) is a new computing paradigm that enables cloud computing and information technology (IT) services to be delivered at the network's edge. By shifting the load of cloud computing to individual local servers, MEC helps meet the requirements of ultralow latency, localized data processing, and extends the potential of Internet of Things (IoT) for end-users. However, the crosscutting nature of MEC and the multidisciplinary components necessary for its deployment have presented additional security and privacy concerns. Fortunately, Artificial Intelligence (AI) algorithms can cope with excessively unpredictable and complex data, which offers a distinct advantage in dealing with sophisticated and developing adversaries in the security industry. Hence, in this paper we comprehensively provide a survey of security and privacy in MEC from the perspective of AI. On the one hand, we use European Telecommunications Standards Institute (ETSI) MEC reference architecture as our based framework while merging the Software Defined Network (SDN) and Network Function Virtualization (NFV) to better illustrate a serviceable platform of MEC. On the other hand, we focus on new security and privacy issues, as well as potential solutions from the viewpoints of AI. Finally, we comprehensively discuss the opportunities and challenges associated with applying AI to MEC security and privacy as possible future research directions.
- Open Access
Backdoor Attacks to Pre-trained Unified Foundation Models
The rise of pre-trained unified foundation models breaks down the barriers between different modalities and tasks, providing comprehensive support to users with unified architectures. However, the backdoor attack on pre-trained models poses a serious threat to their security. Previous research on backdoor attacks has been limited to uni-modal tasks or single tasks across modalities, making it inapplicable to unified foundation models. In this paper, we make proof-of-concept level research on the backdoor attack for pre-trained unified foundation models. Through preliminary experiments on NLP and CV classification tasks, we reveal the vulnerability of these models and suggest future research directions for enhancing the attack approach.
Open AccessBackdoor Attacks to Pre-trained Unified Foundation Models
The rise of pre-trained unified foundation models breaks down the barriers between different modalities and tasks, providing comprehensive support to users with unified architectures. However, the backdoor attack on pre-trained models poses a serious threat to their security. Previous research on backdoor attacks has been limited to uni-modal tasks or single tasks across modalities, making it inapplicable to unified foundation models. In this paper, we make proof-of-concept level research on the backdoor attack for pre-trained unified foundation models. Through preliminary experiments on NLP and CV classification tasks, we reveal the vulnerability of these models and suggest future research directions for enhancing the attack approach.
- Open Access
- 20 April 2022
- Published by Wiley in Mathematical Problems in Engineering
- Vol. 2022, 1-7
Supply Chain Financial Default Risk Early Warning System Based on Particle Swarm Optimization Algorithm
With the advancement of the linkage between financial markets, the probability of credit risk infection is also increasing. Traditional financing methods, which mostly relied on corporate credit to give credit to the whole supply chain, have been replaced by supply chain finance. This paper studies the supply chain financial credit risk through the logistic model and chooses the financial data and supply chain financial operation indicators of relevant listed companies from 2014 to 2016 for analysis. Because not all of companies can find the bad debt rate of accounts receivable from 2014 to 2016, and some agricultural listed companies only have one or two years of relevant data, this paper creates an unbalanced panel data with 91 sample sizes, which is larger than previous studies. Binary logistic regression and principal component analysis are mainly used to accurately calculate the compliance probability of cooperative customers in agricultural supply chain financial products. Unlike the existing literature, which mainly uses s.t to define whether an enterprise defaults, this paper uses value to define the default risk of listed companies in agricultural supply chain finance. In terms of the default risk value of the company, value not only has high accurate value but also has advantages in accurate prediction, which effectively complements and improves the existing research on supply chain finance.
- 20 April 2022
- Published by Wiley in Mathematical Problems in Engineering
- Vol. 2022, 1-7
Open AccessSupply Chain Financial Default Risk Early Warning System Based on Particle Swarm Optimization Algorithm
With the advancement of the linkage between financial markets, the probability of credit risk infection is also increasing. Traditional financing methods, which mostly relied on corporate credit to give credit to the whole supply chain, have been replaced by supply chain finance. This paper studies the supply chain financial credit risk through the logistic model and chooses the financial data and supply chain financial operation indicators of relevant listed companies from 2014 to 2016 for analysis. Because not all of companies can find the bad debt rate of accounts receivable from 2014 to 2016, and some agricultural listed companies only have one or two years of relevant data, this paper creates an unbalanced panel data with 91 sample sizes, which is larger than previous studies. Binary logistic regression and principal component analysis are mainly used to accurately calculate the compliance probability of cooperative customers in agricultural supply chain financial products. Unlike the existing literature, which mainly uses s.t to define whether an enterprise defaults, this paper uses value to define the default risk of listed companies in agricultural supply chain finance. In terms of the default risk value of the company, value not only has high accurate value but also has advantages in accurate prediction, which effectively complements and improves the existing research on supply chain finance.
- RetractedOpen Access
- 27 December 2021
- Published by Wiley in Mobile Information Systems
- Vol. 2021, 1-8
Research on Resource Carbon Emission Allocation Efficiency Based on Blockchain Consensus Algorithm
This paper takes the secondary low-carbon supply chain composed of a manufacturer and a retailer as the research object and takes the government subsidy strategy of low-carbon supply chain as the research theme, the government invests in the construction of blockchain technology application platform, and supply chain enterprises use blockchain technology for compensation. Under this background, considering consumers’ low-carbon preference and green trust, a three-stage game model dominated by the government and followed by manufacturers and retailers under the four scenarios of emission reduction technology input subsidy and low-carbon product output subsidy before and after the application of blockchain technology is established. By comparing the optimal social welfare, carbon emission reduction rate, and low-carbon product output under the four scenarios, the optimal problem of government low-carbon subsidy strategy is discussed. It is found that when the product of consumers’ low-carbon preference coefficient and green trust coefficient is greater than a certain fixed value, the government can obtain higher social welfare and promote enterprises’ emission reduction through output subsidies. The government can always promote the demand for low-carbon products through output subsidies. When the emission reduction cost coefficient approaches infinity, the demand for low-carbon products when the government implements output subsidies are four times that when the government implements technology subsidies. When the single cycle shared cost of the blockchain platform and the unit cost of enterprise application of blockchain technology are less than a certain threshold, the government’s construction of the blockchain platform can obtain higher social welfare, promote enterprise emission reduction, and improve consumers’ demand for low-carbon products.
- 27 December 2021
- Published by Wiley in Mobile Information Systems
- Vol. 2021, 1-8
RetractedOpen AccessResearch on Resource Carbon Emission Allocation Efficiency Based on Blockchain Consensus Algorithm
This paper takes the secondary low-carbon supply chain composed of a manufacturer and a retailer as the research object and takes the government subsidy strategy of low-carbon supply chain as the research theme, the government invests in the construction of blockchain technology application platform, and supply chain enterprises use blockchain technology for compensation. Under this background, considering consumers’ low-carbon preference and green trust, a three-stage game model dominated by the government and followed by manufacturers and retailers under the four scenarios of emission reduction technology input subsidy and low-carbon product output subsidy before and after the application of blockchain technology is established. By comparing the optimal social welfare, carbon emission reduction rate, and low-carbon product output under the four scenarios, the optimal problem of government low-carbon subsidy strategy is discussed. It is found that when the product of consumers’ low-carbon preference coefficient and green trust coefficient is greater than a certain fixed value, the government can obtain higher social welfare and promote enterprises’ emission reduction through output subsidies. The government can always promote the demand for low-carbon products through output subsidies. When the emission reduction cost coefficient approaches infinity, the demand for low-carbon products when the government implements output subsidies are four times that when the government implements technology subsidies. When the single cycle shared cost of the blockchain platform and the unit cost of enterprise application of blockchain technology are less than a certain threshold, the government’s construction of the blockchain platform can obtain higher social welfare, promote enterprise emission reduction, and improve consumers’ demand for low-carbon products.
- RetractedOpen Access
- 1 December 2021
- Published by Wiley in Journal of Mathematics
- Vol. 2021, 1-9
Research on the Risk Measurement Algorithm of Supply Chain Order Financing Based on Insurance Actuarial
In today’s globalized economy, all the links of supply chain are interlinked. Most of the upstream raw material manufacturers or producers in the said chain are small- and medium-sized enterprises (SMEs) that provide the basis for the efficient operation of the whole supply chain. However, SMEs in China, especially those playing a pivotal role in China’s export-oriented economy at this stage, do not have access to the corresponding financial treatment. Supply chain finance provides a new perspective to solve this contradiction. Henceforth, this paper introduces modern financial engineering risk measurement tools to measure the financial risk in supply chain finance, specifically while evaluating the single financing business. Moreover, the chief objective of this paper will be the analysis of the characteristics and connotations of order financing business model. In addition, the focus will be to analyze the risk of order financing from the perspective of banks and other financial institutions. Additionally, this paper will use the CreditRisk + model based on insurance actuarial principles to manage credit risk in order financing business based on foreign currency settlement, in conjunction with the characteristics of supply chain finance and multinational supply chain. Furthermore, a risk measurement method for the application of order financing in multinational supply chains will be provided. Ultimately, the experiments show that the solution of this paper defines and analyzes the financial risks brought by order financing business to bank financing.
- 1 December 2021
- Published by Wiley in Journal of Mathematics
- Vol. 2021, 1-9
RetractedOpen AccessResearch on the Risk Measurement Algorithm of Supply Chain Order Financing Based on Insurance Actuarial
In today’s globalized economy, all the links of supply chain are interlinked. Most of the upstream raw material manufacturers or producers in the said chain are small- and medium-sized enterprises (SMEs) that provide the basis for the efficient operation of the whole supply chain. However, SMEs in China, especially those playing a pivotal role in China’s export-oriented economy at this stage, do not have access to the corresponding financial treatment. Supply chain finance provides a new perspective to solve this contradiction. Henceforth, this paper introduces modern financial engineering risk measurement tools to measure the financial risk in supply chain finance, specifically while evaluating the single financing business. Moreover, the chief objective of this paper will be the analysis of the characteristics and connotations of order financing business model. In addition, the focus will be to analyze the risk of order financing from the perspective of banks and other financial institutions. Additionally, this paper will use the CreditRisk + model based on insurance actuarial principles to manage credit risk in order financing business based on foreign currency settlement, in conjunction with the characteristics of supply chain finance and multinational supply chain. Furthermore, a risk measurement method for the application of order financing in multinational supply chains will be provided. Ultimately, the experiments show that the solution of this paper defines and analyzes the financial risks brought by order financing business to bank financing.