Gaoyang Liu
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
Gaoyang Liu has a total of 55 co-authors. Here are the top co-authors by number of shared publications:
- Chen Wang
Huazhong University of Science and Technology
17 shared publications - Ahmed Mohamed Abdelmoniem
Queen Mary University of London
7 shared publications - Yang Yang
Key Laboratory of Intelligent Sensing System and Security (Ministry of Education), School of Artificial Intelligence, Hubei University, Wuhan, China
4 shared publications - Jiangchuan Liu
Simon Fraser University, Burnaby, Canada
3 shared publications - Kai Peng
Huazhong University of Science and Technology
3 shared publications - Zixiong Wang
Hubei Key Laboratory of Internet of Intelligence, School of EIC, Huazhong University of Science and Technology, Wuhan, China
3 shared publications
Latest Publications
- 1 November 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Dependable and Secure Computing
- Vol. 23 (6) , 12923-12938
Black-Box Membership Inference Attacks Against Contrastive Learning via Aggressive Data Augmentations
- 1 November 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Dependable and Secure Computing
- Vol. 23 (6) , 12923-12938
Black-Box Membership Inference Attacks Against Contrastive Learning via Aggressive Data Augmentations
- Open Access
- 10 August 2026
- Published by Elsevier in SSRN Electronic Journal
Beyond More Agents: A Survey of Collaboration Mechanisms in Multi-Agent LLM Systems
The rapid progress of large language models (LLMs) has enabled agentic AI systems that can reason, plan, use tools, and act with increasing autonomy. However, single-agent systems remain constrained by hallucination, finite context windows, error accumulation, and weak self-verification, particularly in complex or long-horizon tasks. Multi-agent LLM systems offer a promising alternative by assigning complementary roles to LLM agents, enabling task decomposition, parallel execution, cross-agent verification, and the integration of diverse expertise. Yet simply adding more agents does not guarantee better performance than single-agent baselines: its benefits depend on whether improved reasoning and verification can outweigh coordination overhead, error propagation, higher cost, and increased latency. This survey examines collaborative agentic AI through five core dimensions: (D1) coordination, (D2) communication, (D3) negotiation and conflict resolution, (D4) shared knowledge and memory, and (D5) adaptation and learning. Unlike existing surveys that mainly organize the field by application domain, system architecture, or agent role, we focus on how these collaboration mechanisms interact and how they shape system-level performance. This survey first introduces the technical foundations of collaborative agentic AI, explaining how LLMs support agentic behaviors and why interaction among multiple agents can enable capabilities such as distributed reasoning and cross-agent verification that are difficult to achieve within a single agent. We then present a unifying framework around the five dimensions, define each dimension operationally, and discuss where they interact in real systems. For each dimension, we review representative architectures and methods, while highlighting the interactions and trade-offs that arise across dimensions. Finally, we identify open challenges (inc. proactive conflict detection, stable long-horizon adaptation, scalable knowledge and memory governance, and cost-aware collaboration) and outline future directions for building more reliable collaborative agentic AI systems.
- 10 August 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessBeyond More Agents: A Survey of Collaboration Mechanisms in Multi-Agent LLM Systems
The rapid progress of large language models (LLMs) has enabled agentic AI systems that can reason, plan, use tools, and act with increasing autonomy. However, single-agent systems remain constrained by hallucination, finite context windows, error accumulation, and weak self-verification, particularly in complex or long-horizon tasks. Multi-agent LLM systems offer a promising alternative by assigning complementary roles to LLM agents, enabling task decomposition, parallel execution, cross-agent verification, and the integration of diverse expertise. Yet simply adding more agents does not guarantee better performance than single-agent baselines: its benefits depend on whether improved reasoning and verification can outweigh coordination overhead, error propagation, higher cost, and increased latency. This survey examines collaborative agentic AI through five core dimensions: (D1) coordination, (D2) communication, (D3) negotiation and conflict resolution, (D4) shared knowledge and memory, and (D5) adaptation and learning. Unlike existing surveys that mainly organize the field by application domain, system architecture, or agent role, we focus on how these collaboration mechanisms interact and how they shape system-level performance. This survey first introduces the technical foundations of collaborative agentic AI, explaining how LLMs support agentic behaviors and why interaction among multiple agents can enable capabilities such as distributed reasoning and cross-agent verification that are difficult to achieve within a single agent. We then present a unifying framework around the five dimensions, define each dimension operationally, and discuss where they interact in real systems. For each dimension, we review representative architectures and methods, while highlighting the interactions and trade-offs that arise across dimensions. Finally, we identify open challenges (inc. proactive conflict detection, stable long-horizon adaptation, scalable knowledge and memory governance, and cost-aware collaboration) and outline future directions for building more reliable collaborative agentic AI systems.
- Open Access
- 20 April 2026
- Published by Association for Computing Machinery (ACM)
- p. 1739-1750
Breaking the Boundary Barrier: Robust Model Fingerprinting via Unlearnable Examples in Model-Parameter Space
- 20 April 2026
- Published by Association for Computing Machinery (ACM)
- p. 1739-1750
Open AccessBreaking the Boundary Barrier: Robust Model Fingerprinting via Unlearnable Examples in Model-Parameter Space
- Open Access
- 19 November 2025
- Published by Association for Computing Machinery (ACM)
- p. 2804-2817
Prototype Surgery: Tailoring Neural Prototypes via Soft Labels for Efficient Machine Unlearning
- 19 November 2025
- Published by Association for Computing Machinery (ACM)
- p. 2804-2817
Open AccessPrototype Surgery: Tailoring Neural Prototypes via Soft Labels for Efficient Machine Unlearning
- 5 August 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Dependable and Secure Computing
- Vol. 22 (6) , 7256-7271
LLMGraph: Label-Free Detection Against APTs in Edge Networks via LLM and GCN
In the growing trend of remote working, millions of edge networks (e.g., homes or branch offices) are increasingly threatened by Advanced Persistent Threats (APTs), because of the weakened segmentation between business and non-business devices in remote working environment. Despite the fact that numerous APT detection mechanisms have been proposed, all of them are struggling to handle the complex structure, the massive scale and the diverse topology of edge networks. Can recent machine-learning advances tackle these APT detection pain points in edge networks? The GNNs (Graph Neural Networks) seems to be suited to capture the complex structure, but its adjacency matrix fails to capture key network flow context. Additionally, GNNs require extensive manual labeling, which is not scalable. LLMs (Large Language Models) have the potential to provide automatic labeling for the GNNs, but they lack the supplementary security context needed for effective labeling. To address these gaps, we present LLMGraph, which incorporates extended GCNs (Graph Convolutional Networks) and domain-specific RAG (Retrieval-Augmented Generation) pipeline to achieve label-free detection against APTs in edge networks. LLMGraph's extended GCNs model can capture network flow context and direction. LLMGraph's domain-specific RAG pipeline can supplement key security contexts, including device vulnerability and network flow, for effective labeling. Additionally, LLMGraph provides an LLM aggregator to augment and merge the diverse topology of the edge networks. Compared to the state-of-the-art mechanisms, LLMGraph proves effective and scalable, improving the F1-score by at least 46.9%, and the training time for 1 million edge networks is within 1000 s.
- 5 August 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Dependable and Secure Computing
- Vol. 22 (6) , 7256-7271
LLMGraph: Label-Free Detection Against APTs in Edge Networks via LLM and GCN
In the growing trend of remote working, millions of edge networks (e.g., homes or branch offices) are increasingly threatened by Advanced Persistent Threats (APTs), because of the weakened segmentation between business and non-business devices in remote working environment. Despite the fact that numerous APT detection mechanisms have been proposed, all of them are struggling to handle the complex structure, the massive scale and the diverse topology of edge networks. Can recent machine-learning advances tackle these APT detection pain points in edge networks? The GNNs (Graph Neural Networks) seems to be suited to capture the complex structure, but its adjacency matrix fails to capture key network flow context. Additionally, GNNs require extensive manual labeling, which is not scalable. LLMs (Large Language Models) have the potential to provide automatic labeling for the GNNs, but they lack the supplementary security context needed for effective labeling. To address these gaps, we present LLMGraph, which incorporates extended GCNs (Graph Convolutional Networks) and domain-specific RAG (Retrieval-Augmented Generation) pipeline to achieve label-free detection against APTs in edge networks. LLMGraph's extended GCNs model can capture network flow context and direction. LLMGraph's domain-specific RAG pipeline can supplement key security contexts, including device vulnerability and network flow, for effective labeling. Additionally, LLMGraph provides an LLM aggregator to augment and merge the diverse topology of the edge networks. Compared to the state-of-the-art mechanisms, LLMGraph proves effective and scalable, improving the F1-score by at least 46.9%, and the training time for 1 million edge networks is within 1000 s.
- Open Access
- 3 August 2025
- Published by Association for Computing Machinery (ACM)
- p. 215-226
From Expansion to Retraction: Long-tailed Machine Unlearning via Boundary Manipulation
- 3 August 2025
- Published by Association for Computing Machinery (ACM)
- p. 215-226
Open AccessFrom Expansion to Retraction: Long-tailed Machine Unlearning via Boundary Manipulation
- 28 April 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Dependable and Secure Computing
- Vol. 22 (5) , 5277-5290
Poisoning as a Post-Protection: Mitigating Membership Privacy Leakage From Gradient and Prediction of Federated Models
Federated learning (FL) is a distributed learning paradigm that enables multiple clients to train a unified model without sharing their private data. However, recent works demonstrate that FL models are vulnerable to membership inference attacks (MIAs), which can infer whether a data sample was used to train a given FL model. Existing countermeasures either require far-reaching modifications of FL training process or enforce extra processing in prediction phase, yielding them unlikely to be applied well in practice. In this paper, we design a post-protection mechanism, dubbed P$^{2}$ -Protection, which degrades the inference performance of MIAs by simultaneously poisoning the prediction and gradient of the target FL model to reduce the privacy leakage of training data while keeping the model prediction accuracy. P$^{2}$ -Protection only involves one additional training round to embed the poisoned prediction and gradient into the target FL model, without requiring model retraining or training process modification. We evaluate P$^{2}$ -Protection and compare it with two state-of-the-art defenses against three MIAs on five realistic datasets. Experimental results show that P$^{2}$ -Protection outperforms the existing defenses by offering limited implement overhead and improved utility-privacy trade-off.- 28 April 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Dependable and Secure Computing
- Vol. 22 (5) , 5277-5290
Poisoning as a Post-Protection: Mitigating Membership Privacy Leakage From Gradient and Prediction of Federated Models
Federated learning (FL) is a distributed learning paradigm that enables multiple clients to train a unified model without sharing their private data. However, recent works demonstrate that FL models are vulnerable to membership inference attacks (MIAs), which can infer whether a data sample was used to train a given FL model. Existing countermeasures either require far-reaching modifications of FL training process or enforce extra processing in prediction phase, yielding them unlikely to be applied well in practice. In this paper, we design a post-protection mechanism, dubbed P$^{2}$ -Protection, which degrades the inference performance of MIAs by simultaneously poisoning the prediction and gradient of the target FL model to reduce the privacy leakage of training data while keeping the model prediction accuracy. P$^{2}$ -Protection only involves one additional training round to embed the poisoned prediction and gradient into the target FL model, without requiring model retraining or training process modification. We evaluate P$^{2}$ -Protection and compare it with two state-of-the-art defenses against three MIAs on five realistic datasets. Experimental results show that P$^{2}$ -Protection outperforms the existing defenses by offering limited implement overhead and improved utility-privacy trade-off.- 4 April 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Neural Networks and Learning Systems
- Vol. 36 (9) , 15851-15865
Unlearning Attacks for Regression Learning
Recently, the machine unlearning has emerged as a popular method for efficiently erasing the impact of personal data in machine learning (ML) models upon the data owner’s removal request. However, few studies take into consideration the security concerns that may exist in the unlearning process. In this article, we propose the first unlearning attack dubbed unlearning attack for regression learning (UnAR) to deliberately influence the predictive behavior of the target sample against regression learning models. The central concept of UnAR revolves around misleading the regression model into erasing the information associated with the influential samples for the target sample. Observing that the influential samples for target data are generally located far away from the regression plane, we thus propose two novel methods, known as influential sample selection (ISS) and influential sample unlearning (ISU), to identify and subsequently eliminate the lineage of the influential samples. By doing so, we can substantially introduce bias into the prediction pertaining to the target sample, yielding the deliberate manipulation for the user adversely. We extensively evaluate UnAR on five public datasets, and the experimental results indicate our attacks can achieve prediction deviations over 35% by unlearning only 0.5% data as the influential samples.
- 4 April 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Neural Networks and Learning Systems
- Vol. 36 (9) , 15851-15865
Unlearning Attacks for Regression Learning
Recently, the machine unlearning has emerged as a popular method for efficiently erasing the impact of personal data in machine learning (ML) models upon the data owner’s removal request. However, few studies take into consideration the security concerns that may exist in the unlearning process. In this article, we propose the first unlearning attack dubbed unlearning attack for regression learning (UnAR) to deliberately influence the predictive behavior of the target sample against regression learning models. The central concept of UnAR revolves around misleading the regression model into erasing the information associated with the influential samples for the target sample. Observing that the influential samples for target data are generally located far away from the regression plane, we thus propose two novel methods, known as influential sample selection (ISS) and influential sample unlearning (ISU), to identify and subsequently eliminate the lineage of the influential samples. By doing so, we can substantially introduce bias into the prediction pertaining to the target sample, yielding the deliberate manipulation for the user adversely. We extensively evaluate UnAR on five public datasets, and the experimental results indicate our attacks can achieve prediction deviations over 35% by unlearning only 0.5% data as the influential samples.
- Open Access
- 1 January 2025
- Published by Association for Computational Linguistics (ACL)
- p. 13978-13999
Decoupling Memories, Muting Neurons: Towards Practical Machine Unlearning for Large Language Models
Lishuai Hou, Zixiong Wang, Gaoyang Liu, Chen Wang, Wei Liu, Kai Peng. Findings of the Association for Computational Linguistics: ACL 2025. 2025.
- 1 January 2025
- Published by Association for Computational Linguistics (ACL)
- p. 13978-13999
Open AccessDecoupling Memories, Muting Neurons: Towards Practical Machine Unlearning for Large Language Models
Lishuai Hou, Zixiong Wang, Gaoyang Liu, Chen Wang, Wei Liu, Kai Peng. Findings of the Association for Computational Linguistics: ACL 2025. 2025.