Jun Li
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
Jun Li has a total of 52 co-authors. Here are the top co-authors by number of shared publications:
- Guangjun Wu
Chinese Academy of Sciences
7 shared publications Shuqin Zhang
4 shared publications- Shuqin Zhang
School of Computer Science, Zhongyuan University of Technology, ZhengZhou, China
4 shared publications - Fangfang Shan
College of Computer and Artificial Intelligence, Zhongyuan University of Technology, Zhengzhou 451191, China
3 shared publications - Jizhao Liu
College of Computer and Artificial Intelligence, Zhongyuan University of Technology, Zhengzhou 451191, China
3 shared publications Jinbu Geng
2 shared publications
Latest Publications
- 26 September 2026
- Published by Elsevier in Knowledge-Based Systems
- Vol. 352
NIDW: A post-hoc cascade multi-channel text watermarking framework with Newton interpolation
- 26 September 2026
- Published by Elsevier in Knowledge-Based Systems
- Vol. 352
NIDW: A post-hoc cascade multi-channel text watermarking framework with Newton interpolation
- 26 June 2026
- Published by Elsevier in Neurocomputing
- Vol. 699
A hybrid attention-based ACGAN method for intrusion detection with enhanced feature discrimination
- 26 June 2026
- Published by Elsevier in Neurocomputing
- Vol. 699
A hybrid attention-based ACGAN method for intrusion detection with enhanced feature discrimination
- Open Access
- 27 April 2026
- Published by MDPI AG in Information
- Vol. 17 (5) , 416
DualGAD: A Generalist Graph Anomaly Detection Method via Dual-Encoder Architecture
Due to the capability of graph structures to model complex relationships, graph anomaly detection has significant application value in various domains, including financial fraud detection, network security, and fake account identification. Traditional graph anomaly detection methods follow a specialized paradigm of “one dataset, one model”, which requires retraining or fine-tuning models for each new domain. This approach faces critical challenges in practical applications, namely high deployment costs and limited generalization capability. To address this problem, generalist graph anomaly detection aims to achieve the goal of “train once, apply across domains”. However, existing generalist methods primarily rely on graph neural networks to implicitly learn structural information, where the learned structural representations are tightly coupled with specific topology distributions, resulting in limited structural stability under domain shifts. To address this limitation, we propose DualGAD, a generalist graph anomaly detection method via a dual-encoder architecture. In particular, DualGAD introduces explicit structural modeling that characterizes the relative topological deviation of nodes with respect to the overall graph structure, thereby enhancing structural invariance across heterogeneous domains. This method separately models node attribute information and explicit graph structural information via an attribute feature encoder and an explicit structural feature encoder, and adopts an “attribute-dominant, structure-complementary” fusion strategy to achieve collaborative modeling. Experiments on eight real datasets demonstrate that DualGAD achieves an average improvement of 3.12% in AUROC compared to the strongest baseline methods, exhibiting significant cross-domain generalization capability.
- 27 April 2026
- Published by MDPI AG in Information
- Vol. 17 (5) , 416
Open AccessDualGAD: A Generalist Graph Anomaly Detection Method via Dual-Encoder Architecture
Due to the capability of graph structures to model complex relationships, graph anomaly detection has significant application value in various domains, including financial fraud detection, network security, and fake account identification. Traditional graph anomaly detection methods follow a specialized paradigm of “one dataset, one model”, which requires retraining or fine-tuning models for each new domain. This approach faces critical challenges in practical applications, namely high deployment costs and limited generalization capability. To address this problem, generalist graph anomaly detection aims to achieve the goal of “train once, apply across domains”. However, existing generalist methods primarily rely on graph neural networks to implicitly learn structural information, where the learned structural representations are tightly coupled with specific topology distributions, resulting in limited structural stability under domain shifts. To address this limitation, we propose DualGAD, a generalist graph anomaly detection method via a dual-encoder architecture. In particular, DualGAD introduces explicit structural modeling that characterizes the relative topological deviation of nodes with respect to the overall graph structure, thereby enhancing structural invariance across heterogeneous domains. This method separately models node attribute information and explicit graph structural information via an attribute feature encoder and an explicit structural feature encoder, and adopts an “attribute-dominant, structure-complementary” fusion strategy to achieve collaborative modeling. Experiments on eight real datasets demonstrate that DualGAD achieves an average improvement of 3.12% in AUROC compared to the strongest baseline methods, exhibiting significant cross-domain generalization capability.
- Open Access
- 2 March 2026
- Published by MDPI AG in Information
- Vol. 17 (3) , 245
APT-LMSPS: An Efficient APT Detection System via Long-Range Meta-Path Progressive Sampling Search
Advanced Persistent Threats (APTs) are characterized by stealth, infrequency, and long cycles, evading traditional security to endanger critical infrastructure. Complex semantic links between system entities can be accurately modeled using representation learning techniques based on heterogeneous provenance graphs, providing a novel method for uncovering hidden APT attack chains. However, in large-scale practical implementations, this approach still faces three major challenges: combinatorial explosion of long-range meta-paths, loss of semantic evolution during graph compression, and high computational overhead for dynamic environments. To address these, we propose APT-LMSPS, a detection system leveraging Long-Range Meta-path Progressive Sampling Search (LMSPS). The LMSPS algorithm uses dynamic pruning and semantic contribution assessment to convert meta-path combination explosion into constant-scale computation, accurately modeling long-range dependencies. Second, the Maintaining Global Semantics (MGS) approach intelligently filters events by tracking node semantic state changes, achieving an 8:1 compression ratio while preserving over 90% of critical pathways’ semantic integrity. Lastly, the meta-path encoding database uses a caching approach to avoid repeated encoding, doubling encoding effectiveness and enabling efficient, accurate, system-wide APT detection in large-scale scenarios. Evaluated on DARPA, StreamSpot, and ATLAS datasets, APT-LMSPS maintains competitive accuracy (F1-score ≥ 0.98) and improves long-range processing efficiency by an order of magnitude over baselines.
- 2 March 2026
- Published by MDPI AG in Information
- Vol. 17 (3) , 245
Open AccessAPT-LMSPS: An Efficient APT Detection System via Long-Range Meta-Path Progressive Sampling Search
Advanced Persistent Threats (APTs) are characterized by stealth, infrequency, and long cycles, evading traditional security to endanger critical infrastructure. Complex semantic links between system entities can be accurately modeled using representation learning techniques based on heterogeneous provenance graphs, providing a novel method for uncovering hidden APT attack chains. However, in large-scale practical implementations, this approach still faces three major challenges: combinatorial explosion of long-range meta-paths, loss of semantic evolution during graph compression, and high computational overhead for dynamic environments. To address these, we propose APT-LMSPS, a detection system leveraging Long-Range Meta-path Progressive Sampling Search (LMSPS). The LMSPS algorithm uses dynamic pruning and semantic contribution assessment to convert meta-path combination explosion into constant-scale computation, accurately modeling long-range dependencies. Second, the Maintaining Global Semantics (MGS) approach intelligently filters events by tracking node semantic state changes, achieving an 8:1 compression ratio while preserving over 90% of critical pathways’ semantic integrity. Lastly, the meta-path encoding database uses a caching approach to avoid repeated encoding, doubling encoding effectiveness and enabling efficient, accurate, system-wide APT detection in large-scale scenarios. Evaluated on DARPA, StreamSpot, and ATLAS datasets, APT-LMSPS maintains competitive accuracy (F1-score ≥ 0.98) and improves long-range processing efficiency by an order of magnitude over baselines.
- Open Access
- 3 February 2026
- Published by MDPI AG in Mathematics
- Vol. 14 (3) , 551
MAESTRO: A Multi-Scale Ensemble Framework with GAN-Based Data Refinement for Robust Malicious Tor Traffic Detection
Malicious Tor traffic data contains deep domain-specific knowledge, which makes labeling challenging, and the lack of labeled data degrades the accuracy of learning-based detectors. Real-world deployments also exhibit severe class imbalance, where malicious traffic constitutes a small minority of network flows, which further reduces detection performance. In addition, Tor’s fixed 512-byte cell architecture removes packet-size diversity that many encrypted-traffic methods rely on, making feature extraction difficult. This paper proposes an efficient three-stage framework, MAESTRO v1.0, for malicious Tor traffic detection. In Stage 1, MAESTRO extracts multi-scale behavioral signatures by fusing temporal, positional, and directional embeddings at cell, direction, and flow granularities to mitigate feature homogeneity; it then compresses these representations with an autoencoder into compact latent features. In Stage 2, MAESTRO introduces an ensemble-based quality quantification method that combines five complementary anomaly detection models to produce robust discriminability scores for adaptive sample weighting, helping the classifier to emphasize high-quality samples. MAESTRO also trains three specialized GANs per minority class and applies strict five-model ensemble validation to synthesize diverse high-fidelity samples, addressing extreme class imbalance. We evaluate MAESTRO under systematic imbalance settings, ranging from the natural distribution to an extreme 1% malicious ratio. On the CCS’22 Tor malware dataset, MAESTRO achieves 92.38% accuracy, 64.79% recall, and 73.70% F1-score under the natural distribution, improving F1-score by up to 15.53% compared with state-of-the-art baselines. Under the 1% malicious setting, MAESTRO maintains 21.1% recall, which is 14.1 percentage points higher than the best baseline, while conventional methods drop below 10%.
- 3 February 2026
- Published by MDPI AG in Mathematics
- Vol. 14 (3) , 551
Open AccessMAESTRO: A Multi-Scale Ensemble Framework with GAN-Based Data Refinement for Robust Malicious Tor Traffic Detection
Malicious Tor traffic data contains deep domain-specific knowledge, which makes labeling challenging, and the lack of labeled data degrades the accuracy of learning-based detectors. Real-world deployments also exhibit severe class imbalance, where malicious traffic constitutes a small minority of network flows, which further reduces detection performance. In addition, Tor’s fixed 512-byte cell architecture removes packet-size diversity that many encrypted-traffic methods rely on, making feature extraction difficult. This paper proposes an efficient three-stage framework, MAESTRO v1.0, for malicious Tor traffic detection. In Stage 1, MAESTRO extracts multi-scale behavioral signatures by fusing temporal, positional, and directional embeddings at cell, direction, and flow granularities to mitigate feature homogeneity; it then compresses these representations with an autoencoder into compact latent features. In Stage 2, MAESTRO introduces an ensemble-based quality quantification method that combines five complementary anomaly detection models to produce robust discriminability scores for adaptive sample weighting, helping the classifier to emphasize high-quality samples. MAESTRO also trains three specialized GANs per minority class and applies strict five-model ensemble validation to synthesize diverse high-fidelity samples, addressing extreme class imbalance. We evaluate MAESTRO under systematic imbalance settings, ranging from the natural distribution to an extreme 1% malicious ratio. On the CCS’22 Tor malware dataset, MAESTRO achieves 92.38% accuracy, 64.79% recall, and 73.70% F1-score under the natural distribution, improving F1-score by up to 15.53% compared with state-of-the-art baselines. Under the 1% malicious setting, MAESTRO maintains 21.1% recall, which is 14.1 percentage points higher than the best baseline, while conventional methods drop below 10%.
- Open Access
- 2 February 2026
- Published by MDPI AG in Electronics
- Vol. 15 (3) , 628
SCAT: A Spectral-Convolutional Anomaly Transformer for Multivariate Time Series Anomaly Detection
Time series anomaly detection plays a vital role in the supervision of complex systems, including spacecraft operations, industrial production lines, and Internet of Things infrastructures. However, the existing methods face two key challenges: (1) fixed-threshold frequency filters fail to adapt to non-stationary noise, often leading to the loss of critical anomaly signals; and (2) deep models struggle to balance local feature extraction and global temporal dependency, resulting in limited robustness and generalization. To address these problems, we propose the Spectral-Convolutional Anomaly Transformer (SCAT), a unified framework integrating spectral domain adaptive filtering and spatio-temporal gated learning. Specifically, the Spectral Energy Gating Unit (SEGU) dynamically suppresses noise through learnable frequency masking, while Spatio-Temporal Gated Fusion (ST-Gate) combines multi-scale causal convolution and ConvGRU to harmonize local and long-term patterns. A joint optimization strategy further enhances the discrimination between normal and anomalous sequences. Our experiments on five public benchmarks (SMAP, MSL, PSM, SMD, SWaT) showed that SCAT attained an average improvement of 2.46 percentage points on the F1-score relative to leading baseline approaches, demonstrating strong adaptability and robustness in complex noisy environments.
- 2 February 2026
- Published by MDPI AG in Electronics
- Vol. 15 (3) , 628
Open AccessSCAT: A Spectral-Convolutional Anomaly Transformer for Multivariate Time Series Anomaly Detection
Time series anomaly detection plays a vital role in the supervision of complex systems, including spacecraft operations, industrial production lines, and Internet of Things infrastructures. However, the existing methods face two key challenges: (1) fixed-threshold frequency filters fail to adapt to non-stationary noise, often leading to the loss of critical anomaly signals; and (2) deep models struggle to balance local feature extraction and global temporal dependency, resulting in limited robustness and generalization. To address these problems, we propose the Spectral-Convolutional Anomaly Transformer (SCAT), a unified framework integrating spectral domain adaptive filtering and spatio-temporal gated learning. Specifically, the Spectral Energy Gating Unit (SEGU) dynamically suppresses noise through learnable frequency masking, while Spatio-Temporal Gated Fusion (ST-Gate) combines multi-scale causal convolution and ConvGRU to harmonize local and long-term patterns. A joint optimization strategy further enhances the discrimination between normal and anomalous sequences. Our experiments on five public benchmarks (SMAP, MSL, PSM, SMD, SWaT) showed that SCAT attained an average improvement of 2.46 percentage points on the F1-score relative to leading baseline approaches, demonstrating strong adaptability and robustness in complex noisy environments.
- Open Access
- 12 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. PP, 1
Research on Discovery and Mapping of ATT&CK Tactics and Techniques by Cyber Threat Intelligence Based on BERT-TextCNN
Cyber Threat Intelligence (CTI) serves as a critical component in modern cybersecurity. Nevertheless, the different writing styles and lack of standardization in CTI documentation impose significant interpretation overhead on security analysts. To mitigate this challenge, we present a structured methodology for automated extraction and contextual mapping of adversarial tactics and techniques from unstructured threat reports to the ATT&CK framework. Our architecture comprises two core components: a BERT-TextCNN hybrid classifier for feature extraction, and a result correction module that corrects classification results based on the tactics and technique dependencies in the ATT&CK matrix, thereby eliminating errors caused by independent predictions. The model is trained and evaluated on a composite dataset, constructed from the canonical descriptions in the MITRE ATT&CK knowledge base and supplemented with annotated real-world reports from the TRAM (Threat Report ATT&CK Mapping) dataset to enhance generalization. Evaluation results demonstrate classification accuracy of 83.80% for tactics and 71.74% for techniques, with the latter outperforming existing baseline approaches. The proposed method effectively converts unstructured CTI into structured knowledge aligned with ATT&CK, providing a tool to assist security analysts in the intelligence processing pipeline.
- 12 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. PP, 1
Open AccessResearch on Discovery and Mapping of ATT&CK Tactics and Techniques by Cyber Threat Intelligence Based on BERT-TextCNN
Cyber Threat Intelligence (CTI) serves as a critical component in modern cybersecurity. Nevertheless, the different writing styles and lack of standardization in CTI documentation impose significant interpretation overhead on security analysts. To mitigate this challenge, we present a structured methodology for automated extraction and contextual mapping of adversarial tactics and techniques from unstructured threat reports to the ATT&CK framework. Our architecture comprises two core components: a BERT-TextCNN hybrid classifier for feature extraction, and a result correction module that corrects classification results based on the tactics and technique dependencies in the ATT&CK matrix, thereby eliminating errors caused by independent predictions. The model is trained and evaluated on a composite dataset, constructed from the canonical descriptions in the MITRE ATT&CK knowledge base and supplemented with annotated real-world reports from the TRAM (Threat Report ATT&CK Mapping) dataset to enhance generalization. Evaluation results demonstrate classification accuracy of 83.80% for tactics and 71.74% for techniques, with the latter outperforming existing baseline approaches. The proposed method effectively converts unstructured CTI into structured knowledge aligned with ATT&CK, providing a tool to assist security analysts in the intelligence processing pipeline.
- Open Access
- 5 November 2025
- Published by MDPI AG in Information
- Vol. 16 (11) , 959
GSTGPT: A GPT-Based Framework for Multi-Source Data Anomaly Detection
Anomaly detection is a critical approach for ensuring the security of microservice systems. In recent years, deep sequence models have been widely applied to transform sequence modeling into a language modeling problem. However, the objective of training sequence models with language modeling loss is not directly aligned with anomaly detection. Moreover, the diverse data types in microservice systems—namely metrics, logs, and traces—exhibit asynchrony and complex interdependencies. Existing methods based on deep sequence models, such as LogBERT and TranAD, can only account for a limited number of data modalities, failing to fully utilize multi-source data and effectively handle the interrelationships among multiple modalities. To address this, we propose a multimodal anomaly detection framework based on a generative pre-trained language model (GPT), named GSTGPT. GSTGPT represents multi-source data as a feature graph, with metrics and logs as node features and traces as edge features. Additionally, we model feature interactions and dependencies within sequences using spatio-temporal attention and enhance the model’s focus on critical features through feature augmentation. Experimental results on two real-world datasets demonstrate that GSTGPT achieves an F1 score of 0.967, an 8.3% improvement over baseline methods, significantly outperforming them.
- 5 November 2025
- Published by MDPI AG in Information
- Vol. 16 (11) , 959
Open AccessGSTGPT: A GPT-Based Framework for Multi-Source Data Anomaly Detection
Anomaly detection is a critical approach for ensuring the security of microservice systems. In recent years, deep sequence models have been widely applied to transform sequence modeling into a language modeling problem. However, the objective of training sequence models with language modeling loss is not directly aligned with anomaly detection. Moreover, the diverse data types in microservice systems—namely metrics, logs, and traces—exhibit asynchrony and complex interdependencies. Existing methods based on deep sequence models, such as LogBERT and TranAD, can only account for a limited number of data modalities, failing to fully utilize multi-source data and effectively handle the interrelationships among multiple modalities. To address this, we propose a multimodal anomaly detection framework based on a generative pre-trained language model (GPT), named GSTGPT. GSTGPT represents multi-source data as a feature graph, with metrics and logs as node features and traces as edge features. Additionally, we model feature interactions and dependencies within sequences using spatio-temporal attention and enhance the model’s focus on critical features through feature augmentation. Experimental results on two real-world datasets demonstrate that GSTGPT achieves an F1 score of 0.967, an 8.3% improvement over baseline methods, significantly outperforming them.
- 17 September 2025
- Published by Wiley in International Journal of Network Management
- Vol. 35 (6)
A Survey of IPv6 Active Address Discovery
The core of IPv6 active address discovery lies in identifying operational network devices, which serve as the foundation for various network applications. However, the vast address space of IPv6 presents significant challenges to network discovery, including inefficiencies of traditional scanning tools (such as ZMap and Masscan), uncertainties arising from dynamic address allocation, sparsity in address distribution, and limitations of current discovery techniques in regions lacking seed addresses. To address these challenges, researchers have developed various address discovery methods aimed at maximizing the identification of active IPv6 addresses within constrained resource budgets. This paper summarizes and analyzes existing active address discovery methods and evaluates their performance in real‐world environments. Our work includes classifying detection techniques and outlining the current research landscape for address discovery in regions without seed addresses. Through experiments conducted in real network environments, we evaluate and contrast different probing algorithms using four performance metrics: hit rate, hierarchical prefix coverage, aliased addresses, and address discovery rate. Furthermore, we discuss the challenges faced by current IPv6 active address discovery methods in practical applications. Despite numerous advancements, the field of IPv6 active address discovery still requires further research and exploration.
- 17 September 2025
- Published by Wiley in International Journal of Network Management
- Vol. 35 (6)
A Survey of IPv6 Active Address Discovery
The core of IPv6 active address discovery lies in identifying operational network devices, which serve as the foundation for various network applications. However, the vast address space of IPv6 presents significant challenges to network discovery, including inefficiencies of traditional scanning tools (such as ZMap and Masscan), uncertainties arising from dynamic address allocation, sparsity in address distribution, and limitations of current discovery techniques in regions lacking seed addresses. To address these challenges, researchers have developed various address discovery methods aimed at maximizing the identification of active IPv6 addresses within constrained resource budgets. This paper summarizes and analyzes existing active address discovery methods and evaluates their performance in real‐world environments. Our work includes classifying detection techniques and outlining the current research landscape for address discovery in regions without seed addresses. Through experiments conducted in real network environments, we evaluate and contrast different probing algorithms using four performance metrics: hit rate, hierarchical prefix coverage, aliased addresses, and address discovery rate. Furthermore, we discuss the challenges faced by current IPv6 active address discovery methods in practical applications. Despite numerous advancements, the field of IPv6 active address discovery still requires further research and exploration.
- 21 May 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Internet of Things Journal
- Vol. 12 (15) , 30921-30932
A Differential Privacy Based Task Offloading Algorithm for Vehicular Edge Computing
With the advent of vehicular ad hoc networks (VANETs), vehicular edge computing (VEC) facilitates the execution of vehicular tasks through the Internet. In the VEC architecture, vehicles request task offloading, and a central decision center allocates resources. Effective task offloading algorithms provide optimal and equitable decisions based on objectives, such as task latency and system overhead; however, current task offloading algorithms for VEC face challenges in adapting to complex and dynamic road environments. This paper proposes a task-offloading algorithm based on deep reinforcement learning (RL) to address the challenges of task offloading in VEC. During the task offloading process, the privacy of vehicular task data may be compromised. This study introduces a novel task-offloading algorithm for VEC that employs differential privacy (DP) principles to safeguard the confidentiality of vehicular tasks during the offloading process. The proposed algorithm introduces noise in accordance with the privacy budget during the training process. The study provides a theoretical analysis of privacy, and experimental results based on Attari demonstrate that the proposed DP-based RL algorithm exhibits superior convergence compared to existing algorithms. Veins-based simulation experiments on VEC demonstrate that the proposed DP-based task-offloading algorithm can achieve practical offloading while preserving privacy.
- 21 May 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Internet of Things Journal
- Vol. 12 (15) , 30921-30932
A Differential Privacy Based Task Offloading Algorithm for Vehicular Edge Computing
With the advent of vehicular ad hoc networks (VANETs), vehicular edge computing (VEC) facilitates the execution of vehicular tasks through the Internet. In the VEC architecture, vehicles request task offloading, and a central decision center allocates resources. Effective task offloading algorithms provide optimal and equitable decisions based on objectives, such as task latency and system overhead; however, current task offloading algorithms for VEC face challenges in adapting to complex and dynamic road environments. This paper proposes a task-offloading algorithm based on deep reinforcement learning (RL) to address the challenges of task offloading in VEC. During the task offloading process, the privacy of vehicular task data may be compromised. This study introduces a novel task-offloading algorithm for VEC that employs differential privacy (DP) principles to safeguard the confidentiality of vehicular tasks during the offloading process. The proposed algorithm introduces noise in accordance with the privacy budget during the training process. The study provides a theoretical analysis of privacy, and experimental results based on Attari demonstrate that the proposed DP-based RL algorithm exhibits superior convergence compared to existing algorithms. Veins-based simulation experiments on VEC demonstrate that the proposed DP-based task-offloading algorithm can achieve practical offloading while preserving privacy.