Ming Hu
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Ming Hu has a total of 149 co-authors. Here are the top co-authors by number of shared publications:
- Mingsong Chen
East China Normal University, Shanghai, China
25 shared publications - Xiaofei Xie
Singapore Management University, Singapore, Singapore
10 shared publications Mingsong Chen
8 shared publications- Yang Liu
Department of Obstetrics and Gynecology, The Second Affiliated Hospital of Kunming Medical University, China
7 shared publications Jun Xia
6 shared publicationsXiaofei Xie
School of Computing and Information Systems, Singapore Management University, Victoria Street, Singapore
6 shared publications
Latest Publications
- Open Access
- 9 November 2026
- Published by Association for Computing Machinery (ACM)
- p. 125-134
Suit the Remedy to the Case: Precise Backdoor Elimination for Text-to-Image Diffusion Models via Model Editing
With the widespread adoption of Text-to-Image (T2I) Diffusion Models (DMs) across various applications, their security has attracted increasing attention. In particular, these models are highly vulnerable to backdoor attacks, where adversaries implant hidden triggers during training to manipulate model outputs at inference time. To deal with the backdoor attacks, existing defense methods typically require substantial training data and employ coarse-grained alignment, leading to catastrophic forgetting and semantic degradation, respectively. To defend against backdoor attacks while preventing performance degradation, this paper first analyzes the behavior of poisoned models and finds that i) the neuron activation variation distributions of trigger tokens are significantly different from those of normal tokens, and ii) the embeddings of non-semantic trigger tokens are far from the manifold space constructed by all the normal tokens in the vocabulary. Based on the findings, we propose ManiEdit, a novel model-edit-based backdoor elimination framework for T2I DMs. Specifically, we propose an activation-based module localization methodology to pinpoint the backdoored module by calculating the Neuron Activation Variations (NAV), enabling precise parameter recoupling within restricted areas to safeguard benign parameters from collateral damage. To sever malicious associations and restore the original semantics of triggers, we present a manifold-driven trigger categorization strategy, which detects triggers via NAV within the localized modules, subsequently applying syntactic and geometric constraints to distinguish non-semantic from semantic triggers. By aligning non-semantic triggers with empty prompts and realigning semantic triggers with denoising-optimized semantic anchors, ManiEdit can achieve efficient elimination without expensive retraining. Comprehensive experiments demonstrate the effectiveness of ManiEdit in defending against various state-of-the-art backdoor attack methods.
- 9 November 2026
- Published by Association for Computing Machinery (ACM)
- p. 125-134
Open AccessSuit the Remedy to the Case: Precise Backdoor Elimination for Text-to-Image Diffusion Models via Model Editing
With the widespread adoption of Text-to-Image (T2I) Diffusion Models (DMs) across various applications, their security has attracted increasing attention. In particular, these models are highly vulnerable to backdoor attacks, where adversaries implant hidden triggers during training to manipulate model outputs at inference time. To deal with the backdoor attacks, existing defense methods typically require substantial training data and employ coarse-grained alignment, leading to catastrophic forgetting and semantic degradation, respectively. To defend against backdoor attacks while preventing performance degradation, this paper first analyzes the behavior of poisoned models and finds that i) the neuron activation variation distributions of trigger tokens are significantly different from those of normal tokens, and ii) the embeddings of non-semantic trigger tokens are far from the manifold space constructed by all the normal tokens in the vocabulary. Based on the findings, we propose ManiEdit, a novel model-edit-based backdoor elimination framework for T2I DMs. Specifically, we propose an activation-based module localization methodology to pinpoint the backdoored module by calculating the Neuron Activation Variations (NAV), enabling precise parameter recoupling within restricted areas to safeguard benign parameters from collateral damage. To sever malicious associations and restore the original semantics of triggers, we present a manifold-driven trigger categorization strategy, which detects triggers via NAV within the localized modules, subsequently applying syntactic and geometric constraints to distinguish non-semantic from semantic triggers. By aligning non-semantic triggers with empty prompts and realigning semantic triggers with denoising-optimized semantic anchors, ManiEdit can achieve efficient elimination without expensive retraining. Comprehensive experiments demonstrate the effectiveness of ManiEdit in defending against various state-of-the-art backdoor attack methods.
- Open Access
- 7 October 2026
- Published by Association for Computing Machinery (ACM)
- p. 916-928
Clarity Is Not Assumed: Understanding LLM-Based Code Generation under Ambiguous Requirements
Software requirement ambiguity is ubiquitous in real-world development, stemming from the inherent imprecision of natural language and the varying interpretations of stakeholders. While Large Language Models (LLMs) have demonstrated impressive capabilities in generating code from precise specifications, such ambiguity poses a significant obstacle to reliable automated code generation. Existing benchmarks typically assume clear and unambiguous requirements, leaving an empirical gap in understanding how LLMs behave when faced with the inherent uncertainty of real-world software requirements. In this paper, we introduce Orchid, a code generation benchmark specifically designed with ambiguous requirements. It comprises 1,304 function-level tasks covering four distinct types of ambiguity: lexical, syntactic, semantic, and vagueness. Leveraging this dataset, we conduct the first systematic empirical study to evaluate the impact of requirement ambiguity on LLM-based code generation. Our results demonstrate that ambiguity consistently degrades the performance of all evaluated LLMs, with the most pronounced negative effects observed in highly advanced models. Furthermore, we observe that LLMs frequently produce functionally divergent implementations for the same ambiguous requirement and lack the capability to identify or resolve such ambiguity autonomously. These findings reveal a significant performance gap between clear and ambiguous requirements, underscoring the urgent need for ambiguity-aware techniques in the next generation of automated software engineering tools. The Orchid benchmark is publicly available at https://huggingface.co/datasets/SII-YDD/Orchid.
- 7 October 2026
- Published by Association for Computing Machinery (ACM)
- p. 916-928
Open AccessClarity Is Not Assumed: Understanding LLM-Based Code Generation under Ambiguous Requirements
Software requirement ambiguity is ubiquitous in real-world development, stemming from the inherent imprecision of natural language and the varying interpretations of stakeholders. While Large Language Models (LLMs) have demonstrated impressive capabilities in generating code from precise specifications, such ambiguity poses a significant obstacle to reliable automated code generation. Existing benchmarks typically assume clear and unambiguous requirements, leaving an empirical gap in understanding how LLMs behave when faced with the inherent uncertainty of real-world software requirements. In this paper, we introduce Orchid, a code generation benchmark specifically designed with ambiguous requirements. It comprises 1,304 function-level tasks covering four distinct types of ambiguity: lexical, syntactic, semantic, and vagueness. Leveraging this dataset, we conduct the first systematic empirical study to evaluate the impact of requirement ambiguity on LLM-based code generation. Our results demonstrate that ambiguity consistently degrades the performance of all evaluated LLMs, with the most pronounced negative effects observed in highly advanced models. Furthermore, we observe that LLMs frequently produce functionally divergent implementations for the same ambiguous requirement and lack the capability to identify or resolve such ambiguity autonomously. These findings reveal a significant performance gap between clear and ambiguous requirements, underscoring the urgent need for ambiguity-aware techniques in the next generation of automated software engineering tools. The Orchid benchmark is publicly available at https://huggingface.co/datasets/SII-YDD/Orchid.
- Open Access
- 1 October 2026
- Published by Association for Computing Machinery (ACM) in Proceedings of the ACM on Software Engineering
- Vol. 3 (ISSTA) , 4765-4788
Silence of Commit Messages: An Empirical Study for Vulnerability Commit Message Generation using Large Language Models
The vulnerability commit message serves as crucial metadata for maintaining software within version control systems. Nonetheless, manually crafted vulnerability commit messages often lack detail or exhibit inconsistent formatting. Recently, the growing use of Large Language Models (LLMs) for code and natural language comprehension has opened avenues to automate the crafting of these messages. This paper systematically and thoroughly explores the generation of security patch commit messages in the context of LLMs, delving into topics such as dataset construction, evaluation method design, and the relationship between vulnerability types and submission structure. First, we explore the elements of commit messages using LLMs and integrate a questionnaire survey to pinpoint four essential types of information: summary, background, impact, and fix, which are essential for developers. This aims to establish a structured dataset of bug submissions and assess its quality. Next, we examine general automated evaluation techniques for assessing LLM-generated commit messages and find that GPT-3.5's evaluation methods align more closely with human judgment. Then, we conduct an organized investigation into how LLM generation effects vary across three principal vulnerability types, uncovering that LLMs' adaptability differs across vulnerabilities. Furthermore, we perform an exhaustive examination of generation quality across various components and find that LLMs excel at generating summaries but struggle to produce impact details. In particular, the smallest DeepSeek-Coder shows a semantic retention advantage in crafting backgrounds, whereas DeepSeek-V3 struggles with impact aspects. Lastly, we investigate the effects of different prompting strategies (e.g., zero-shot, few-shot prompts) and parameter settings (e.g., temperature and top_p) on the quality of commit message generation, finding that prompt and parameter configurations critically influence output quality, with model sensitivity varying.
- 1 October 2026
- Published by Association for Computing Machinery (ACM) in Proceedings of the ACM on Software Engineering
- Vol. 3 (ISSTA) , 4765-4788
Open AccessSilence of Commit Messages: An Empirical Study for Vulnerability Commit Message Generation using Large Language Models
The vulnerability commit message serves as crucial metadata for maintaining software within version control systems. Nonetheless, manually crafted vulnerability commit messages often lack detail or exhibit inconsistent formatting. Recently, the growing use of Large Language Models (LLMs) for code and natural language comprehension has opened avenues to automate the crafting of these messages. This paper systematically and thoroughly explores the generation of security patch commit messages in the context of LLMs, delving into topics such as dataset construction, evaluation method design, and the relationship between vulnerability types and submission structure. First, we explore the elements of commit messages using LLMs and integrate a questionnaire survey to pinpoint four essential types of information: summary, background, impact, and fix, which are essential for developers. This aims to establish a structured dataset of bug submissions and assess its quality. Next, we examine general automated evaluation techniques for assessing LLM-generated commit messages and find that GPT-3.5's evaluation methods align more closely with human judgment. Then, we conduct an organized investigation into how LLM generation effects vary across three principal vulnerability types, uncovering that LLMs' adaptability differs across vulnerabilities. Furthermore, we perform an exhaustive examination of generation quality across various components and find that LLMs excel at generating summaries but struggle to produce impact details. In particular, the smallest DeepSeek-Coder shows a semantic retention advantage in crafting backgrounds, whereas DeepSeek-V3 struggles with impact aspects. Lastly, we investigate the effects of different prompting strategies (e.g., zero-shot, few-shot prompts) and parameter settings (e.g., temperature and top_p) on the quality of commit message generation, finding that prompt and parameter configurations critically influence output quality, with model sensitivity varying.
- 1 October 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Internet of Things Journal
- Vol. 13 (20) , 47915-47929
Information Scenario-Aware Incentive Mechanisms for Federated Learning Using Game Theory and Reinforcement Learning
- 1 October 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Internet of Things Journal
- Vol. 13 (20) , 47915-47929
Information Scenario-Aware Incentive Mechanisms for Federated Learning Using Game Theory and Reinforcement Learning
- Open Access
- 8 August 2026
- Published by Association for Computing Machinery (ACM)
- p. 13439-13440
FedKDD/FedMAS 2026: The 2026 International Joint Workshop on Federated Learning for Multi-agent Systems and Data Mining
Multi-agent systems (MAS) are enabling increasingly complex, collaborative applications in autonomous driving, smart logistics, robotic coordination, and distributed sensing. Their effectiveness depends on collective intelligence emerging from multiple distributed agents, each operating with partial information and often sensitive local data. To realize such collaboration while preserving data privacy and autonomy, federated learning (FL) has emerged as a de facto decentralized framework that allows agents to learn shared models without centralizing raw data. Despite the rapid progress in both FL and MAS, significant challenges remain in integrating these paradigms—such as coordinating heterogeneous agents, handling non-IID data, ensuring communication efficiency, and maintaining system robustness and fairness in open environments. This workshop seeks to bring together researchers and practitioners from academia and industry to explore the convergence of federated learning and multi-agent systems. We aim to foster discussions on foundational advances, real-world deployments, and emerging interdisciplinary opportunities, with a focus on scalability, trustworthiness, adaptive coordination, and the broader societal impact of federated multi-agent intelligence.
- 8 August 2026
- Published by Association for Computing Machinery (ACM)
- p. 13439-13440
Open AccessFedKDD/FedMAS 2026: The 2026 International Joint Workshop on Federated Learning for Multi-agent Systems and Data Mining
Multi-agent systems (MAS) are enabling increasingly complex, collaborative applications in autonomous driving, smart logistics, robotic coordination, and distributed sensing. Their effectiveness depends on collective intelligence emerging from multiple distributed agents, each operating with partial information and often sensitive local data. To realize such collaboration while preserving data privacy and autonomy, federated learning (FL) has emerged as a de facto decentralized framework that allows agents to learn shared models without centralizing raw data. Despite the rapid progress in both FL and MAS, significant challenges remain in integrating these paradigms—such as coordinating heterogeneous agents, handling non-IID data, ensuring communication efficiency, and maintaining system robustness and fairness in open environments. This workshop seeks to bring together researchers and practitioners from academia and industry to explore the convergence of federated learning and multi-agent systems. We aim to foster discussions on foundational advances, real-world deployments, and emerging interdisciplinary opportunities, with a focus on scalability, trustworthiness, adaptive coordination, and the broader societal impact of federated multi-agent intelligence.
- Open Access
- 20 April 2026
- Published by Association for Computing Machinery (ACM)
- p. 59-70
Two Heads Are Better Than One: Generalized Cross-Domain Federated Learning via Dual-Prototype
- 20 April 2026
- Published by Association for Computing Machinery (ACM)
- p. 59-70
Open AccessTwo Heads Are Better Than One: Generalized Cross-Domain Federated Learning via Dual-Prototype
- Open Access
- 12 April 2026
- Published by Association for Computing Machinery (ACM)
- p. 2735-2746
Unlocking the Silent Needs: Business-Logic-Driven Iterative Requirements Auto-completion
- 12 April 2026
- Published by Association for Computing Machinery (ACM)
- p. 2735-2746
Open AccessUnlocking the Silent Needs: Business-Logic-Driven Iterative Requirements Auto-completion
- Open Access
- 12 April 2026
- Published by Association for Computing Machinery (ACM)
- p. 2786-2797
Modeling Like Peeling an Onion: Layerwise Analysis-Driven Automatic Behavioral Model Generation
As software complexity skyrockets and requirements evolve at breakneck speed, traditional human-centric behavioral modeling can no longer keep pace in terms of efficiency, accuracy, and scalability. While existing automated approaches can produce models, they still struggle with deep semantic understanding of textual requirements or with reasoning about intricate system logic, especially nested relationships. Inspired by the way experienced analysts “peel back” layers of a problem, we propose LATO, a Layerwise Analysis-Driven AuTomatic Behavioral MOdeling approach. It employs a progressive decomposition strategy to guide large language models in incrementally parsing requirement structures, deconstructing behavioral dependencies, and ultimately generating executable UML activity diagrams. Comprehensive evaluations on four open-source datasets and two real-world industrial systems show that LATO comprehensively outperforms state-of-the-art baselines in accuracy, completeness, and syntactic compliance: F1 scores for behavioral node-extraction improve by up to 71.1%, relation-extraction F1 by 52.4% relatively, and syntactic pass rates remain above 96.67%. The framework also exhibits strong robustness to input perturbations, confirming its cross-domain generalizability. This paper is the first to tightly fuse human-inspired strategies with LLMs in behavior modeling, yielding an intelligent infrastructure that exhibits expert-level logical understanding and generalization. By closing the modeling-skills gap, LATO delivers a next-generation, low-cost, and explainable solution for requirements engineering and AI-native software development.
- 12 April 2026
- Published by Association for Computing Machinery (ACM)
- p. 2786-2797
Open AccessModeling Like Peeling an Onion: Layerwise Analysis-Driven Automatic Behavioral Model Generation
As software complexity skyrockets and requirements evolve at breakneck speed, traditional human-centric behavioral modeling can no longer keep pace in terms of efficiency, accuracy, and scalability. While existing automated approaches can produce models, they still struggle with deep semantic understanding of textual requirements or with reasoning about intricate system logic, especially nested relationships. Inspired by the way experienced analysts “peel back” layers of a problem, we propose LATO, a Layerwise Analysis-Driven AuTomatic Behavioral MOdeling approach. It employs a progressive decomposition strategy to guide large language models in incrementally parsing requirement structures, deconstructing behavioral dependencies, and ultimately generating executable UML activity diagrams. Comprehensive evaluations on four open-source datasets and two real-world industrial systems show that LATO comprehensively outperforms state-of-the-art baselines in accuracy, completeness, and syntactic compliance: F1 scores for behavioral node-extraction improve by up to 71.1%, relation-extraction F1 by 52.4% relatively, and syntactic pass rates remain above 96.67%. The framework also exhibits strong robustness to input perturbations, confirming its cross-domain generalizability. This paper is the first to tightly fuse human-inspired strategies with LLMs in behavior modeling, yielding an intelligent infrastructure that exhibits expert-level logical understanding and generalization. By closing the modeling-skills gap, LATO delivers a next-generation, low-cost, and explainable solution for requirements engineering and AI-native software development.
- Open Access
- 16 February 2026
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Software Engineering and Methodology
- Vol. 35 (10) , 1-27
Separate Bombs from Firecrackers: Intention-guided Vulnerability Assessment via LLM
Although modern vulnerability detection tools enable developers to efficiently identify numerous security flaws, indiscriminate remediation efforts often result in unnecessary development expenses. This is particularly true given that a substantial portion of detected vulnerabilities are either lowly exploitable or have a negligible impact in practical operational environments. Consequently, vulnerability severity assessment has emerged as a critical component in optimizing software development efficiency. Existing vulnerability assessment methods typically rely on manually crafted descriptions associated with source code artifacts. However, due to variability in description quality and subjectivity in intention interpretation, the performance of these methods is seriously limited. To address this issue, this paper introduces VulStamp, a novel intention-guided framework that facilitates description-free vulnerability assessment. Specifically, VulStamp employs static analysis in conjunction with a Large Language Model (LLM) to extract the intention information from vulnerable code. Based on the intention information, VulStamp employs a prompt-tuned model for vulnerability assessment. Furthermore, to mitigate the problem of imbalanced data associated with vulnerability types, VulStamp integrates a Reinforcement Learning (RL)-based prompt-tuning method to train the assessment model. Extensive experimental results demonstrate that VulStamp outperforms the state-of-the-art baselines by an average of 12.9%, 102.6%, 18.3%, and 54.1% in terms of AUC, precision, recall, and F1-score, respectively.
- 16 February 2026
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Software Engineering and Methodology
- Vol. 35 (10) , 1-27
Open AccessSeparate Bombs from Firecrackers: Intention-guided Vulnerability Assessment via LLM
Although modern vulnerability detection tools enable developers to efficiently identify numerous security flaws, indiscriminate remediation efforts often result in unnecessary development expenses. This is particularly true given that a substantial portion of detected vulnerabilities are either lowly exploitable or have a negligible impact in practical operational environments. Consequently, vulnerability severity assessment has emerged as a critical component in optimizing software development efficiency. Existing vulnerability assessment methods typically rely on manually crafted descriptions associated with source code artifacts. However, due to variability in description quality and subjectivity in intention interpretation, the performance of these methods is seriously limited. To address this issue, this paper introduces VulStamp, a novel intention-guided framework that facilitates description-free vulnerability assessment. Specifically, VulStamp employs static analysis in conjunction with a Large Language Model (LLM) to extract the intention information from vulnerable code. Based on the intention information, VulStamp employs a prompt-tuned model for vulnerability assessment. Furthermore, to mitigate the problem of imbalanced data associated with vulnerability types, VulStamp integrates a Reinforcement Learning (RL)-based prompt-tuning method to train the assessment model. Extensive experimental results demonstrate that VulStamp outperforms the state-of-the-art baselines by an average of 12.9%, 102.6%, 18.3%, and 54.1% in terms of AUC, precision, recall, and F1-score, respectively.
- Top Cited PapersOpen Access
- 11 December 2025
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Software Engineering and Methodology
- Vol. 35 (1) , 1-40
JailGuard : A Universal Detection Framework for Prompt-based Attacks on LLM Systems
The systems and software powered by Large Language Models (LLMs) and Multi-Modal Large Language Models (MLLMs) have played a critical role in numerous scenarios. However, current LLM systems are vulnerable to prompt-based attacks, with jailbreaking attacks enabling the LLM system to generate harmful content, while hijacking attacks manipulate the LLM system to perform attacker-desired tasks, underscoring the necessity for detection tools. Unfortunately, existing detecting approaches are usually tailored to specific attacks, resulting in poor generalization in detecting various attacks across different modalities. To address it, we propose JailGuard , a universal detection framework deployed on top of LLM systems for prompt-based attacks across text and image modalities. JailGuard operates on the principle that attacks are inherently less robust than benign ones. Specifically, JailGuard mutates untrusted inputs to generate variants and leverages the discrepancy of the variants’ responses on the target model to distinguish attack samples from benign samples. We implement 18 mutators for text and image inputs and design a mutator combination policy to further improve detection generalization. The evaluation on the dataset containing 15 known attack types suggests that JailGuard achieves the best detection accuracy of 86.14%/82.90% on text and image inputs, outperforming state-of-the-art methods by 11.81–25.73% and 12.20–21.40%.
- 11 December 2025
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Software Engineering and Methodology
- Vol. 35 (1) , 1-40
Top Cited PapersOpen AccessJailGuard : A Universal Detection Framework for Prompt-based Attacks on LLM Systems
The systems and software powered by Large Language Models (LLMs) and Multi-Modal Large Language Models (MLLMs) have played a critical role in numerous scenarios. However, current LLM systems are vulnerable to prompt-based attacks, with jailbreaking attacks enabling the LLM system to generate harmful content, while hijacking attacks manipulate the LLM system to perform attacker-desired tasks, underscoring the necessity for detection tools. Unfortunately, existing detecting approaches are usually tailored to specific attacks, resulting in poor generalization in detecting various attacks across different modalities. To address it, we propose JailGuard , a universal detection framework deployed on top of LLM systems for prompt-based attacks across text and image modalities. JailGuard operates on the principle that attacks are inherently less robust than benign ones. Specifically, JailGuard mutates untrusted inputs to generate variants and leverages the discrepancy of the variants’ responses on the target model to distinguish attack samples from benign samples. We implement 18 mutators for text and image inputs and design a mutator combination policy to further improve detection generalization. The evaluation on the dataset containing 15 known attack types suggests that JailGuard achieves the best detection accuracy of 86.14%/82.90% on text and image inputs, outperforming state-of-the-art methods by 11.81–25.73% and 12.20–21.40%.