Ruyun Zhang
Research Subjects
Ruyun Zhang has a total of 72 co-authors. Here are the top co-authors by number of shared publications:
- Chao Zha
Institute of Computing Technology, Chinese Academy of Sciences, Beijing, China
6 shared publications - Bing Bai
Research Center for High Efficiency Computing Infrastructure, Zhejiang Lab, Hangzhou, Zhejiang, China
4 shared publications Gang Zhao
National Key Laboratory of Science and Technology on Information System Security
4 shared publications- Shunbin Li
Zhejiang Lab, Hangzhou, Zhejiang, China
4 shared publications Xiaohui Kuang
National Key Laboratory of Science and Technology on Information System Security
4 shared publications- Zhiyu Wang
Intelligent Computing Infrastructure Innovation Center, Zhejiang Lab, Hangzhou, Zhejiang, China
3 shared publications
Latest Publications
- Open Access
- 14 November 2026
- Published by Association for Computing Machinery (ACM)
- p. 4537-4551
Navigating the Latent Manifold: Proactive Concept Drift Adaptation for Resilient NIDS
- 14 November 2026
- Published by Association for Computing Machinery (ACM)
- p. 4537-4551
Open AccessNavigating the Latent Manifold: Proactive Concept Drift Adaptation for Resilient NIDS
- 29 April 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Dependable and Secure Computing
- Vol. 23 (4) , 8829-8844
Normality in Anomaly: Rethinking Traffic Labels
Network traffic detection plays a critical role in protecting digital infrastructures from increasingly sophisticated cyber threats. Although AI-based detection methods have shown clear advantages over traditional rule-based approaches, most existing approaches either rely on the unrealistic closed-set assumption or address distribution shifts only through model retraining after data drift. As a result, the robustness of detection models under fixed closed-set training conditions remains insufficiently explored. To address this limitation, we revisit malicious traffic modeling from an attacker's perspective and formulate the detection process as a Min–Max adversarial optimization problem. Instead of relying on conventional binary labels, we introduce a self-generating mask mechanism with sparsity constraints, which dynamically identifies malicious feature components and produces adaptive labels during training. This design allows the model to focus on attack-relevant feature dimensions while suppressing benign patterns embedded within malicious traffic. The proposed method is architecture-agnostic and can be integrated with multiple neural architectures, including multilayer perceptrons and Transformers. Extensive ablation studies and comparative evaluations conducted in a real-world environment and three public datasets, CICIDS-2017, UNSW-NB15, and Bot-IoT, demonstrate that the proposed approach consistently outperforms existing baselines in terms of detection robustness and classification performance, achieving improvements of up to 138% in accuracy, 65% in the F1 score and 52% in TPR.
- 29 April 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Dependable and Secure Computing
- Vol. 23 (4) , 8829-8844
Normality in Anomaly: Rethinking Traffic Labels
Network traffic detection plays a critical role in protecting digital infrastructures from increasingly sophisticated cyber threats. Although AI-based detection methods have shown clear advantages over traditional rule-based approaches, most existing approaches either rely on the unrealistic closed-set assumption or address distribution shifts only through model retraining after data drift. As a result, the robustness of detection models under fixed closed-set training conditions remains insufficiently explored. To address this limitation, we revisit malicious traffic modeling from an attacker's perspective and formulate the detection process as a Min–Max adversarial optimization problem. Instead of relying on conventional binary labels, we introduce a self-generating mask mechanism with sparsity constraints, which dynamically identifies malicious feature components and produces adaptive labels during training. This design allows the model to focus on attack-relevant feature dimensions while suppressing benign patterns embedded within malicious traffic. The proposed method is architecture-agnostic and can be integrated with multiple neural architectures, including multilayer perceptrons and Transformers. Extensive ablation studies and comparative evaluations conducted in a real-world environment and three public datasets, CICIDS-2017, UNSW-NB15, and Bot-IoT, demonstrate that the proposed approach consistently outperforms existing baselines in terms of detection robustness and classification performance, achieving improvements of up to 138% in accuracy, 65% in the F1 score and 52% in TPR.
- 27 February 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Cognitive Communications and Networking
- Vol. 12, 6524-6538
Sparse Gaussian–Markov Modeling for Robust and Trustworthy Unknown Cyber Defense
Network intrusion detection systems (NIDS) are essential for enhancing network security by detecting and responding to attacks promptly. NIDS based on artificial intelligence (AI) offer better adaptability to unknown (out-of-distribution (OOD)) attacks compared to traditional anomaly detection methods. However, many existing AI-based NIDS directly adopt methods from other domains without adequately considering OOD data, which often leads to severe overfitting in trained classifiers and insufficient model reliability and robustness. This issue largely arises because most approaches focus solely on the relationship between individual features and predictions while neglecting the interactions among features, resulting in overfitting and overconfident predictions. Moreover, the retention of certain weak parameters may cause the model to maintain artificially high confidence when confronted with OOD data. To address this, we propose a Gaussian Markov Random Field that better captures traffic data distribution and feature correlations. Furthermore, our GMRF-based classification model uses sparsification with regularization $L_{0}$ to remove weak parameters. $L_{0}$ regularization counts the number of non-zero parameters and penalizes model complexity, thereby forcing insignificant weights to become exactly zero and improving generalization. This prevents high prediction probabilities for unknown attacks and effectively distinguishes them from benign traffic. Finally, we evaluated our method in the CICIDS-2017 and CSE-CICIDS-2018 datasets, achieving recall rates of 57% and 77% for unknown attacks while maintaining a recall of more than 95% for known traffic. Our approach significantly outperforms existing SOTA methods, and performance tests on a low-specification machine show a low detection latency of approximately 130 microseconds, making it suitable for real-time applications.
- 27 February 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Cognitive Communications and Networking
- Vol. 12, 6524-6538
Sparse Gaussian–Markov Modeling for Robust and Trustworthy Unknown Cyber Defense
Network intrusion detection systems (NIDS) are essential for enhancing network security by detecting and responding to attacks promptly. NIDS based on artificial intelligence (AI) offer better adaptability to unknown (out-of-distribution (OOD)) attacks compared to traditional anomaly detection methods. However, many existing AI-based NIDS directly adopt methods from other domains without adequately considering OOD data, which often leads to severe overfitting in trained classifiers and insufficient model reliability and robustness. This issue largely arises because most approaches focus solely on the relationship between individual features and predictions while neglecting the interactions among features, resulting in overfitting and overconfident predictions. Moreover, the retention of certain weak parameters may cause the model to maintain artificially high confidence when confronted with OOD data. To address this, we propose a Gaussian Markov Random Field that better captures traffic data distribution and feature correlations. Furthermore, our GMRF-based classification model uses sparsification with regularization $L_{0}$ to remove weak parameters. $L_{0}$ regularization counts the number of non-zero parameters and penalizes model complexity, thereby forcing insignificant weights to become exactly zero and improving generalization. This prevents high prediction probabilities for unknown attacks and effectively distinguishes them from benign traffic. Finally, we evaluated our method in the CICIDS-2017 and CSE-CICIDS-2018 datasets, achieving recall rates of 57% and 77% for unknown attacks while maintaining a recall of more than 95% for known traffic. Our approach significantly outperforms existing SOTA methods, and performance tests on a low-specification machine show a low detection latency of approximately 130 microseconds, making it suitable for real-time applications.
- 12 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Mobile Computing
- Vol. 25 (6) , 9315-9331
FlowXpert: Context-Aware Flow Embedding for Enhanced Traffic Detection in IoT Network
In the Internet of Things (IoT) environment, continuous interaction among a large number of devices generates complex and dynamic network traffic, which poses significant challenges to rule-based detection approaches. Machine learning (ML)-based traffic detection technology, capable of identifying anomalous patterns and potential threats within this traffic, serves as a critical component in ensuring network security. This study first identifies a significant issue with widely adopted feature extraction tools (e.g., CICFlowMeter): the extensive use of time- and length-related features leads to high sparsity, which adversely affects model convergence. Furthermore, existing traffic detection methods generally lack an embedding mechanism capable of efficiently and comprehensively capturing the semantic characteristics of network traffic. To address these challenges, we propose a novel feature extraction tool that eliminates traditional time and length features in favor of context-aware semantic features related to the source host, thus improving the generalizability of the model. In addition, we design an embedding training framework that integrates the unsupervised DBSCAN clustering algorithm with a contrastive learning strategy to effectively capture fine-grained semantic representations of traffic. Extensive empirical evaluations are conducted on the real-world dataset (Mawi) and two simulated datasets (CICIDS-2017 and UNSW-NB15) to validate the proposed method in terms of detection accuracy, robustness, and generalization. Comparative experiments against several state-of-the-art (SOTA) models demonstrate the superior performance of our approach. Furthermore, we confirm its applicability and deployability in real-time scenarios.
- 12 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Mobile Computing
- Vol. 25 (6) , 9315-9331
FlowXpert: Context-Aware Flow Embedding for Enhanced Traffic Detection in IoT Network
In the Internet of Things (IoT) environment, continuous interaction among a large number of devices generates complex and dynamic network traffic, which poses significant challenges to rule-based detection approaches. Machine learning (ML)-based traffic detection technology, capable of identifying anomalous patterns and potential threats within this traffic, serves as a critical component in ensuring network security. This study first identifies a significant issue with widely adopted feature extraction tools (e.g., CICFlowMeter): the extensive use of time- and length-related features leads to high sparsity, which adversely affects model convergence. Furthermore, existing traffic detection methods generally lack an embedding mechanism capable of efficiently and comprehensively capturing the semantic characteristics of network traffic. To address these challenges, we propose a novel feature extraction tool that eliminates traditional time and length features in favor of context-aware semantic features related to the source host, thus improving the generalizability of the model. In addition, we design an embedding training framework that integrates the unsupervised DBSCAN clustering algorithm with a contrastive learning strategy to effectively capture fine-grained semantic representations of traffic. Extensive empirical evaluations are conducted on the real-world dataset (Mawi) and two simulated datasets (CICIDS-2017 and UNSW-NB15) to validate the proposed method in terms of detection accuracy, robustness, and generalization. Comparative experiments against several state-of-the-art (SOTA) models demonstrate the superior performance of our approach. Furthermore, we confirm its applicability and deployability in real-time scenarios.
- 29 April 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Network and Service Management
- Vol. 22 (4) , 3646-3661
DM-IDS—A Network Intrusion Detection Method Based on Dual-Modal Fusion
The machine learning-based approach to network intrusion detection presents a groundbreaking research paradigm, positioned to replace traditional rule-based and signature-based methods. However, prior research methodologies have predominantly focused on flow-based approaches, which may not be effective in detecting all types of attacks at a granular level. In this study, we introduce DM-IDS, an attention-convolution architecture model for bimodal network intrusion detection in both flow and payload modalities, using bilinear fusion. Notably, we present a novel method for constructing binary-form feature vectors under the payload modality, with the goal of extracting additional security semantic features. To facilitate this, we independently develop a feature generation tool named Beeman. Finally, we conduct a series of comparative and ablation experiments on two publicly available datasets, CICIDS-2017 and CICIoT-2023, achieving state-of-the-art model performance.
- 29 April 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Network and Service Management
- Vol. 22 (4) , 3646-3661
DM-IDS—A Network Intrusion Detection Method Based on Dual-Modal Fusion
The machine learning-based approach to network intrusion detection presents a groundbreaking research paradigm, positioned to replace traditional rule-based and signature-based methods. However, prior research methodologies have predominantly focused on flow-based approaches, which may not be effective in detecting all types of attacks at a granular level. In this study, we introduce DM-IDS, an attention-convolution architecture model for bimodal network intrusion detection in both flow and payload modalities, using bilinear fusion. Notably, we present a novel method for constructing binary-form feature vectors under the payload modality, with the goal of extracting additional security semantic features. To facilitate this, we independently develop a feature generation tool named Beeman. Finally, we conduct a series of comparative and ablation experiments on two publicly available datasets, CICIDS-2017 and CICIoT-2023, achieving state-of-the-art model performance.
- 11 April 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Internet of Things Journal
- Vol. 12 (13) , 25849-25866
Intelligent Intrusion Detection System With Autonomous Optimal Traffic Steering for Aerial-Aided Edge Computing
Aerial-aided edge computing (AEC) promises to provide low-latency computing services as a critical component for future low-altitude intelligent transportation systems. However, the complex edge network environment poses security challenges, as traditional intrusion detection systems (IDSs) struggle to handle AEC’s large traffic volume and resource constraints. To address this, we propose a collaborative detection mechanism called Switch IDS. First, Switch IDS adopts a packet-level detection solution to meet real-time detection requirements. Next, Switch IDS is designed for resource-constrained nodes. By introducing the Mixture of Experts (MoE) structure into the traditional deep learning (DL) model, it enables seamless and scalable multinode deployment. Switch IDS establishes a resource status-based capability for each Expert and incorporates steering loss in the loss function to enable autonomous near-optimal traffic steering, ultimately maximizing system processing capacity. Finally, the Switch IDS utilizes parallelized service function chaining (SFC) for practical multinode deployment. To the best of our knowledge, this is the first realization of multinode deployment and autonomous traffic steering for DL-based IDS. Experiments on public datasets show that Switch IDS and its multinode deployment scheme notably boost processing capacity while maintaining high detection performance.
- 11 April 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Internet of Things Journal
- Vol. 12 (13) , 25849-25866
Intelligent Intrusion Detection System With Autonomous Optimal Traffic Steering for Aerial-Aided Edge Computing
Aerial-aided edge computing (AEC) promises to provide low-latency computing services as a critical component for future low-altitude intelligent transportation systems. However, the complex edge network environment poses security challenges, as traditional intrusion detection systems (IDSs) struggle to handle AEC’s large traffic volume and resource constraints. To address this, we propose a collaborative detection mechanism called Switch IDS. First, Switch IDS adopts a packet-level detection solution to meet real-time detection requirements. Next, Switch IDS is designed for resource-constrained nodes. By introducing the Mixture of Experts (MoE) structure into the traditional deep learning (DL) model, it enables seamless and scalable multinode deployment. Switch IDS establishes a resource status-based capability for each Expert and incorporates steering loss in the loss function to enable autonomous near-optimal traffic steering, ultimately maximizing system processing capacity. Finally, the Switch IDS utilizes parallelized service function chaining (SFC) for practical multinode deployment. To the best of our knowledge, this is the first realization of multinode deployment and autonomous traffic steering for DL-based IDS. Experiments on public datasets show that Switch IDS and its multinode deployment scheme notably boost processing capacity while maintaining high detection performance.
- 14 March 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Information Forensics and Security
- Vol. 20 (15566013) , 3204-3219
A-NIDS: Adaptive Network Intrusion Detection System Based on Clustering and Stacked CTGAN
Intrusion detection systems (IDS) are crucial tools for detecting anomalous network traffic in cybersecurity. In recent years, significant progress has been made in applying artificial intelligence to IDS. However, existing research often assumes that training and testing data are static and identically distributed, whereas in reality, data drift is inevitable. Moreover, to enhance model versatility and detection performance, models have become increasingly complex, posing challenges to real-time deployment. To address these challenges, we propose an adaptive network intrusion detection system named A-NIDS, consisting of a main task and two bypass tasks. The main task is to develop a fully connected and shallow network with strong detection performance and real-time capability. The first bypass task is a clustering model that helps the main task detect data drift in an unsupervised manner. The second bypass task is a generation model to generate old data to address catastrophic forgetting in new model iterations and the storage cost issue caused by accumulating old data. We conduct extensive experiments on the CICIDS-2017 and CSE-CICIDS-2018 datasets, demonstrating the superior performance of A-NIDS on new and old data. Furthermore, our detection module achieves a detection latency of 5 microseconds, highlighting its suitability for real-time applications. All the related code is publicly available at: https://github.com/ids-sec-hub/A-NIDS.
- 14 March 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Information Forensics and Security
- Vol. 20 (15566013) , 3204-3219
A-NIDS: Adaptive Network Intrusion Detection System Based on Clustering and Stacked CTGAN
Intrusion detection systems (IDS) are crucial tools for detecting anomalous network traffic in cybersecurity. In recent years, significant progress has been made in applying artificial intelligence to IDS. However, existing research often assumes that training and testing data are static and identically distributed, whereas in reality, data drift is inevitable. Moreover, to enhance model versatility and detection performance, models have become increasingly complex, posing challenges to real-time deployment. To address these challenges, we propose an adaptive network intrusion detection system named A-NIDS, consisting of a main task and two bypass tasks. The main task is to develop a fully connected and shallow network with strong detection performance and real-time capability. The first bypass task is a clustering model that helps the main task detect data drift in an unsupervised manner. The second bypass task is a generation model to generate old data to address catastrophic forgetting in new model iterations and the storage cost issue caused by accumulating old data. We conduct extensive experiments on the CICIDS-2017 and CSE-CICIDS-2018 datasets, demonstrating the superior performance of A-NIDS on new and old data. Furthermore, our detection module achieves a detection latency of 5 microseconds, highlighting its suitability for real-time applications. All the related code is publicly available at: https://github.com/ids-sec-hub/A-NIDS.
- Open Access
- 17 September 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Very Large Scale Integration (VLSI) Systems
- Vol. 33 (2) , 512-524
Architectural Exploration for Waferscale Switching System
With the end of Moore’s law and Dennard scaling, waferscale systems or processors that integrate multiple pre-tested known good dies (KGDs) on a waferscale-interposer are new approaches to further improve the chiplet-based system’s performance. This article explores the network on wafer (NoW) architecture of waferscale switching system under several physical constraints. A software-based approach is proposed to redefine the topological property. A five-level butterfly fat-tree (BFT)-like logical topology with 8.96-Tb/s (896 ports $\times 10$ Gb/s/port) switching bandwidth is achieved based on 2-D-mesh-like physical topology. We show that the proposed BFT-like topology with breadth-first-search (BFS) based traffic balanced routing algorithm reduces 55.6% hops, 41.4% transmission delay, and improves 24.2% throughput compared to 2-D-mesh-like topology under different traffic distributions. This BFT-like waferscale switching system is suitable for high-performance computing and data centers. In addition, the numerical analysis shows that the waferscale package can provide significant power efficiency and latency advantages compared to the typical single-chip package, which mainly benefits from the short-reach IO requirements. Note that the proposed waferscale switching system is compatible with high-switch-capacity dies with advanced process technology, which can further improve system performance. Finally, we present the physical implementations for the waferscale system with heterogeneous dies.
- 17 September 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Very Large Scale Integration (VLSI) Systems
- Vol. 33 (2) , 512-524
Open AccessArchitectural Exploration for Waferscale Switching System
With the end of Moore’s law and Dennard scaling, waferscale systems or processors that integrate multiple pre-tested known good dies (KGDs) on a waferscale-interposer are new approaches to further improve the chiplet-based system’s performance. This article explores the network on wafer (NoW) architecture of waferscale switching system under several physical constraints. A software-based approach is proposed to redefine the topological property. A five-level butterfly fat-tree (BFT)-like logical topology with 8.96-Tb/s (896 ports $\times 10$ Gb/s/port) switching bandwidth is achieved based on 2-D-mesh-like physical topology. We show that the proposed BFT-like topology with breadth-first-search (BFS) based traffic balanced routing algorithm reduces 55.6% hops, 41.4% transmission delay, and improves 24.2% throughput compared to 2-D-mesh-like topology under different traffic distributions. This BFT-like waferscale switching system is suitable for high-performance computing and data centers. In addition, the numerical analysis shows that the waferscale package can provide significant power efficiency and latency advantages compared to the typical single-chip package, which mainly benefits from the short-reach IO requirements. Note that the proposed waferscale switching system is compatible with high-switch-capacity dies with advanced process technology, which can further improve system performance. Finally, we present the physical implementations for the waferscale system with heterogeneous dies.
- Open Access
- 4 August 2024
- Published by Association for Computing Machinery (ACM)
- p. 41-47
CollaSFC: An Intelligent Collaborative Approach for In-network SFC Failure Detection in Data Center for AI Computing
The successful application cases of Large Language Models (LLMs) and Machine Learning (ML) are driving traditional data centers to transform into intelligent computing data centers characterized by low latency, high bandwidth, high reliability, and zero packet loss. The demand for immense computing and ultra-low latency suggests that in-network computing (INC) may be a viable solution, such as In-network aggregation (INA). INA involves a hierarchical structure of switches and servers to form different Service Function Chains (SFCs) including switches, servers, physical links, and virtual links for accomplishing model training. However, the aggregation of heavy traffic in CTCs tends to a sudden and drastic increase in a specific node, greatly increasing the likelihood of node failure. To detect SFC failure in real time, we propose an in-network SFC failure detection approach based on INC. We introduce digital twins (DT) and propose a collaborative AI framework based on the data plane and control plane to avoid model overfitting. In addition, to reduce the computing consumption, we propose the concept of "multiple SFC chains multiple models" to customize each SFC failure detection model and validate the mechanism on a BMv2-based prototype, which implements a high-accuracy failure detection with minor performance degradation.
- 4 August 2024
- Published by Association for Computing Machinery (ACM)
- p. 41-47
Open AccessCollaSFC: An Intelligent Collaborative Approach for In-network SFC Failure Detection in Data Center for AI Computing
The successful application cases of Large Language Models (LLMs) and Machine Learning (ML) are driving traditional data centers to transform into intelligent computing data centers characterized by low latency, high bandwidth, high reliability, and zero packet loss. The demand for immense computing and ultra-low latency suggests that in-network computing (INC) may be a viable solution, such as In-network aggregation (INA). INA involves a hierarchical structure of switches and servers to form different Service Function Chains (SFCs) including switches, servers, physical links, and virtual links for accomplishing model training. However, the aggregation of heavy traffic in CTCs tends to a sudden and drastic increase in a specific node, greatly increasing the likelihood of node failure. To detect SFC failure in real time, we propose an in-network SFC failure detection approach based on INC. We introduce digital twins (DT) and propose a collaborative AI framework based on the data plane and control plane to avoid model overfitting. In addition, to reduce the computing consumption, we propose the concept of "multiple SFC chains multiple models" to customize each SFC failure detection model and validate the mechanism on a BMv2-based prototype, which implements a high-accuracy failure detection with minor performance degradation.
- 17 November 2022
- Published by Springer Nature
- p. 389-400
An Efficient Soft Analytical Side-Channel Attack on Ascon
- 17 November 2022
- Published by Springer Nature
- p. 389-400
An Efficient Soft Analytical Side-Channel Attack on Ascon