Wenhan Ge
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Wenhan Ge has a total of 24 co-authors. Here are the top co-authors by number of shared publications:
- Junfeng Wang
College of Computer Science, Sichuan University, China
7 shared publications - Xiaohui Li
School of Cyber Science and Engineering, Sichuan University, Chengdu, China
3 shared publications - Zhiyang Fang
Sichuan University
3 shared publications - Binhui Tang
School of Cyber Science and Engineering, Sichuan University, Chengdu, China
2 shared publications Junfeng Wang
2 shared publications- Xiaolan Zhu
National Key Laboratory of Fundamental Science on Synthetic Vision, Sichuan University, Chengdu, China
2 shared publications
Latest Publications
- 19 May 2026
- Published by Springer Nature
- p. 22-40
IncreTTP: An Incremental TTPs Classification Model
- 20 April 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Information Forensics and Security
- Vol. PP (15566013) , 1
ThreatMAMBA: Achieving High-Robustness Cyber Threat Attribution During the Evolution of Attacks
Cyber Threat Attribution (CTA) involves identifying the perpetrator of a cyberattack by analyzing evidence from attack incidents. As the final step in attributing cybercrime, it plays a critical role in accountability and incident characterization. To improve identification accuracy, most existing CTA methods rely on deep learning and machine learning techniques, integrating heterogeneous knowledge graphs from diverse information sources to support criminal reasoning. However, for ongoing attacks, these post-hoc learned models may not provide intuitive or actionable investigative leads during the early stages of an incident, limiting their applicability in real-time CTA scenarios. To address this challenge, we propose ThreatMAMBA, a temporally robust reasoning and explanation framework for CTA. The method first constructs a heterogeneous information graph by extracting Indicators of Compromise (IOCs), Tactics, Techniques, and Procedures (TTPs), and temporal relationships from Cyber Threat Intelligence (CTI). It then encodes the knowledge graph by preserving semantically significant states using a state space selection mechanism. Finally, contrastive learning is introduced to partition and construct temporal knowledge graphs according to the chronological development of the incident, enabling accurate attacker identification. Experiments show that at event progression stages of 20%, 40%, 60%, 80%, and 100%, ThreatMAMBA improves the Macro F1 score by 6.94%, 5.09%, 9.33%, 10.70%, and 12.05%. This means ThreatMAMBA is a robust and reliable attribution during the early stage of attacks. Moreover, ThreatMAMBA can identify TTPs that positively correlate with specific attacker groups. The attacker behavior profiles generated by our system achieve a Jaccard similarity of 0.2991 with MITRE ATT&CK, outperforming Qi’anxin by 0.1836.
- 20 April 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Information Forensics and Security
- Vol. PP (15566013) , 1
ThreatMAMBA: Achieving High-Robustness Cyber Threat Attribution During the Evolution of Attacks
Cyber Threat Attribution (CTA) involves identifying the perpetrator of a cyberattack by analyzing evidence from attack incidents. As the final step in attributing cybercrime, it plays a critical role in accountability and incident characterization. To improve identification accuracy, most existing CTA methods rely on deep learning and machine learning techniques, integrating heterogeneous knowledge graphs from diverse information sources to support criminal reasoning. However, for ongoing attacks, these post-hoc learned models may not provide intuitive or actionable investigative leads during the early stages of an incident, limiting their applicability in real-time CTA scenarios. To address this challenge, we propose ThreatMAMBA, a temporally robust reasoning and explanation framework for CTA. The method first constructs a heterogeneous information graph by extracting Indicators of Compromise (IOCs), Tactics, Techniques, and Procedures (TTPs), and temporal relationships from Cyber Threat Intelligence (CTI). It then encodes the knowledge graph by preserving semantically significant states using a state space selection mechanism. Finally, contrastive learning is introduced to partition and construct temporal knowledge graphs according to the chronological development of the incident, enabling accurate attacker identification. Experiments show that at event progression stages of 20%, 40%, 60%, 80%, and 100%, ThreatMAMBA improves the Macro F1 score by 6.94%, 5.09%, 9.33%, 10.70%, and 12.05%. This means ThreatMAMBA is a robust and reliable attribution during the early stage of attacks. Moreover, ThreatMAMBA can identify TTPs that positively correlate with specific attacker groups. The attacker behavior profiles generated by our system achieve a Jaccard similarity of 0.2991 with MITRE ATT&CK, outperforming Qi’anxin by 0.1836.
- 7 April 2026
- Published by Elsevier in Applied Soft Computing
- Vol. 197
OSEAF: A robust one-shot environment-aware framework for malware detection evasion
- 7 April 2026
- Published by Elsevier in Applied Soft Computing
- Vol. 197
OSEAF: A robust one-shot environment-aware framework for malware detection evasion
- 12 February 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Information Forensics and Security
- Vol. 21 (15566013) , 2220-2235
MT-DEGCL: Multi-Task Encrypted Traffic Classification With Dual Embedding and Graph Contrastive Learning
Although encryption offers strong anonymity, it also facilitates the concealment of malicious activities, allowing adversaries to evade detection, and posing a great challenge to cybersecurity surveillance. Many existing encrypted traffic classification methods struggle to integrate flow- and packet-level tasks effectively, as they are trained independently, which is redundancy. Additionally, packet header and payload are treated equally, leading to the rich information in raw bytes remains fully unexplored, particularly in the abundant payload data. Moreover, they neglect the semantic invariance and common features between data samples, which ultimately results in suboptimal performance. To address these challenges, we propose an effective Multi-Task model using Dual Embedding and Graph Contrastive Learning (MT-DEGCL). Based on the byte-packet-flow structure of network traffic, a parallel dual embedding embeds the header and payload separately, followed by a cross-gated feature fusion strategy to capture the strong local packet-level representation. Then, we construct the traffic interaction graph and further utilize graph contrastive learning to extract the robust global flow-level representation. Finally, a multi-task model is trained for joint flow- and packet-level classification, leveraging the complementary learning between tasks to enhance overall performance. The experimental results on four real datasets highlight the effectiveness of MT-DEGCL, demonstrating superior performance in both tasks. Specifically, on the ISCX-Tor dataset, MT-DEGCL achieves F1 scores of 98.63% for flow-level classification and 98.10% at the packet level, surpassing the state-of-the-art (i.e., DE-GNN) by 2.03% and 83.21%, respectively. Furthermore, MT-DEGCL maximizes the rich information in raw payload bytes, significantly reducing or even nearly eliminating classification loss when using only payload data.
- 12 February 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Information Forensics and Security
- Vol. 21 (15566013) , 2220-2235
MT-DEGCL: Multi-Task Encrypted Traffic Classification With Dual Embedding and Graph Contrastive Learning
Although encryption offers strong anonymity, it also facilitates the concealment of malicious activities, allowing adversaries to evade detection, and posing a great challenge to cybersecurity surveillance. Many existing encrypted traffic classification methods struggle to integrate flow- and packet-level tasks effectively, as they are trained independently, which is redundancy. Additionally, packet header and payload are treated equally, leading to the rich information in raw bytes remains fully unexplored, particularly in the abundant payload data. Moreover, they neglect the semantic invariance and common features between data samples, which ultimately results in suboptimal performance. To address these challenges, we propose an effective Multi-Task model using Dual Embedding and Graph Contrastive Learning (MT-DEGCL). Based on the byte-packet-flow structure of network traffic, a parallel dual embedding embeds the header and payload separately, followed by a cross-gated feature fusion strategy to capture the strong local packet-level representation. Then, we construct the traffic interaction graph and further utilize graph contrastive learning to extract the robust global flow-level representation. Finally, a multi-task model is trained for joint flow- and packet-level classification, leveraging the complementary learning between tasks to enhance overall performance. The experimental results on four real datasets highlight the effectiveness of MT-DEGCL, demonstrating superior performance in both tasks. Specifically, on the ISCX-Tor dataset, MT-DEGCL achieves F1 scores of 98.63% for flow-level classification and 98.10% at the packet level, surpassing the state-of-the-art (i.e., DE-GNN) by 2.03% and 83.21%, respectively. Furthermore, MT-DEGCL maximizes the rich information in raw payload bytes, significantly reducing or even nearly eliminating classification loss when using only payload data.
- 16 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Internet of Things Journal
- Vol. 13 (7) , 13472-13485
LiteJam: A Lightweight Deep Learning Architecture for Real-Time GNSS Interference Detection and Characterization in UAVs
Global navigation satellite system (GNSS) interference poses a serious threat to uncrewed aerial vehicles (UAVs), potentially leading to navigation failures, airspace violations, or even loss of flight control. Although deep learning methods have demonstrated strong performance in interference detection and characterization, most models remain too computationally expensive for onboard deployment due to their high computational cost. To address this challenge, we propose LiteJam, a lightweight architecture that utilizes precorrelation in-phase and quadrature (I/Q) data to construct pseudo-image representations without requiring additional hardware. Specifically, LiteJam adopts a multiscale convolutional architecture to capture interference patterns, employs a dynamic sparse attention (DSA) mechanism to adaptively emphasize spatiospectral cues, and leverages a hierarchical multihead module for interference detection and characterization. Experimental results show that LiteJam outperforms all baselines. The $F1$ -score of interference classification is 95.74%, outperforming lightweight baselines by 4.37%–24.51%, and generalizes well across diverse scenarios, while maintaining high computational efficiency for real-time UAV applications. Our codes and models are available at https://github.com/CynthiaCYX/LiteJam
- 16 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Internet of Things Journal
- Vol. 13 (7) , 13472-13485
LiteJam: A Lightweight Deep Learning Architecture for Real-Time GNSS Interference Detection and Characterization in UAVs
Global navigation satellite system (GNSS) interference poses a serious threat to uncrewed aerial vehicles (UAVs), potentially leading to navigation failures, airspace violations, or even loss of flight control. Although deep learning methods have demonstrated strong performance in interference detection and characterization, most models remain too computationally expensive for onboard deployment due to their high computational cost. To address this challenge, we propose LiteJam, a lightweight architecture that utilizes precorrelation in-phase and quadrature (I/Q) data to construct pseudo-image representations without requiring additional hardware. Specifically, LiteJam adopts a multiscale convolutional architecture to capture interference patterns, employs a dynamic sparse attention (DSA) mechanism to adaptively emphasize spatiospectral cues, and leverages a hierarchical multihead module for interference detection and characterization. Experimental results show that LiteJam outperforms all baselines. The $F1$ -score of interference classification is 95.74%, outperforming lightweight baselines by 4.37%–24.51%, and generalizes well across diverse scenarios, while maintaining high computational efficiency for real-time UAV applications. Our codes and models are available at https://github.com/CynthiaCYX/LiteJam
- 18 June 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Information Forensics and Security
- Vol. 20 (15566013) , 6340-6355
WF-TFC: An Open-World Few-Shot Anonymous Website Fingerprinting via Time-Frequency Consistency
While Tor provides strong anonymity, it also facilitates the concealment of malicious activities, which poses a significant challenge to cybersecurity surveillance. As an effective anti-anonymity technique, Website Fingerprinting(WF) enables the inference of which websites a user is visiting, thereby uncovering potential attacker activities. State-of-the-art(SOTA) methods have demonstrated remarkable effectiveness. However, a large number of labeled traffic is required to ensure effectiveness, and without timely updates, these models will encounter serious challenges of concept drift due to the dynamic nature of website content and network conditions. The core reasons lie in the independently and identically distributed assumption, while in challenging open-world scenarios, the long-term spatial and temporal dynamics complicates data consistency and effective knowledge transfer. To address these issues, this paper presents WF-TFC, an open-world few-shot anonymous WF model via self-supervised contrastive learning and time-frequency consistency. It aligns time- and frequency-based representations in the latent time-frequency space, enhancing the sustained effectiveness of inherent patterns across various websites. Consequently, it accommodates diverse few-shot target domains with varying dynamics, facilitating data consistency and knowledge transfer in unobserved long-term temporal and spatial environments. For instance, with only 5 traces per website, WF-TFC achieves 92.62% accuracy on traces collected six weeks after pre-training, exceeding the SOTA(i.e., NetCLR) by 2.12%. On similar but mutually exclusive traces, it attains an F1 score of 87.20%, surpassing the SOTA by 6.12%.
- 18 June 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Information Forensics and Security
- Vol. 20 (15566013) , 6340-6355
WF-TFC: An Open-World Few-Shot Anonymous Website Fingerprinting via Time-Frequency Consistency
While Tor provides strong anonymity, it also facilitates the concealment of malicious activities, which poses a significant challenge to cybersecurity surveillance. As an effective anti-anonymity technique, Website Fingerprinting(WF) enables the inference of which websites a user is visiting, thereby uncovering potential attacker activities. State-of-the-art(SOTA) methods have demonstrated remarkable effectiveness. However, a large number of labeled traffic is required to ensure effectiveness, and without timely updates, these models will encounter serious challenges of concept drift due to the dynamic nature of website content and network conditions. The core reasons lie in the independently and identically distributed assumption, while in challenging open-world scenarios, the long-term spatial and temporal dynamics complicates data consistency and effective knowledge transfer. To address these issues, this paper presents WF-TFC, an open-world few-shot anonymous WF model via self-supervised contrastive learning and time-frequency consistency. It aligns time- and frequency-based representations in the latent time-frequency space, enhancing the sustained effectiveness of inherent patterns across various websites. Consequently, it accommodates diverse few-shot target domains with varying dynamics, facilitating data consistency and knowledge transfer in unobserved long-term temporal and spatial environments. For instance, with only 5 traces per website, WF-TFC achieves 92.62% accuracy on traces collected six weeks after pre-training, exceeding the SOTA(i.e., NetCLR) by 2.12%. On similar but mutually exclusive traces, it attains an F1 score of 87.20%, surpassing the SOTA by 6.12%.
- 16 April 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Information Forensics and Security
- Vol. 20 (15566013) , 4915-4928
CorreFlow: A Covert Fingerprinting Modulation for Flow Correlation in Open Heterogeneous Networks
The constantly changing landscape of the Internet presents a significant challenge in the detection and tracking of covert attackers and their sophisticated methods. To address this issue, various techniques, such as network flow watermarking (NFW) and traffic correlation, embed attack labels in data streams to identify attack pathways or aid post-analysis. However, existing solutions are often tailored to specific scenarios, resulting in lacking robustness, adaptability, and anonymity under non-cooperative or incomplete information heterogeneous environments. To this end, this paper proposes CorreFlow, a Transfer Learning (TL) based invisible network flow correlation framework utilizing time channel graph fingerprinting modulation. It considers the fragmentation and reassembly of data packets during transmission. In simple network environments, CorreFlow utilizes TL for rapid correlation across flows, enabling efficient linkage of related traffic segments. In complex heterogeneous network, where traditional correlation methods may fail due to encryption and variability, it leverages Inter Packet Delay (IPD) for encrypted flow matching and accurately identifies the optimal watermark point. Multiple experiments conducted on real network traffic and public datasets have demonstrated that CorreFlow achieves highly efficient traffic correlation with minimal false positive rate, improved adaptability, and steganography. Specifically, it has achieved over 97.31% accuracy in various network environments and promotes network traffic correlation in open heterogeneous network environments from low correlation to 95%.
- 16 April 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Information Forensics and Security
- Vol. 20 (15566013) , 4915-4928
CorreFlow: A Covert Fingerprinting Modulation for Flow Correlation in Open Heterogeneous Networks
The constantly changing landscape of the Internet presents a significant challenge in the detection and tracking of covert attackers and their sophisticated methods. To address this issue, various techniques, such as network flow watermarking (NFW) and traffic correlation, embed attack labels in data streams to identify attack pathways or aid post-analysis. However, existing solutions are often tailored to specific scenarios, resulting in lacking robustness, adaptability, and anonymity under non-cooperative or incomplete information heterogeneous environments. To this end, this paper proposes CorreFlow, a Transfer Learning (TL) based invisible network flow correlation framework utilizing time channel graph fingerprinting modulation. It considers the fragmentation and reassembly of data packets during transmission. In simple network environments, CorreFlow utilizes TL for rapid correlation across flows, enabling efficient linkage of related traffic segments. In complex heterogeneous network, where traditional correlation methods may fail due to encryption and variability, it leverages Inter Packet Delay (IPD) for encrypted flow matching and accurately identifies the optimal watermark point. Multiple experiments conducted on real network traffic and public datasets have demonstrated that CorreFlow achieves highly efficient traffic correlation with minimal false positive rate, improved adaptability, and steganography. Specifically, it has achieved over 97.31% accuracy in various network environments and promotes network traffic correlation in open heterogeneous network environments from low correlation to 95%.
- 4 March 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Information Forensics and Security
- Vol. 19 (15566013) , 3829-3843
MetaCluster: A Universal Interpretable Classification Framework for Cybersecurity
Rising cyber threats have created an immediate demand for Deep Learning (DL) in cybersecurity. Nevertheless, the opaque nature of DL models poses challenges in deploying, collaborating, and assessing their effectiveness in less reliable cybersecurity environments. Despite eXplainable Artificial Intelligence (XAI) playing a role in enhancing cybersecurity analytics, the limited task scope, the propensity for data overfitting, and the stochastic explanations hinder its broader application. To fill the gap, this paper introduces a generic interpretable classification framework, named MetaCluster. MetaCluster generates semantic prototypes for features, patterns, and domains at varying granular levels by following three fundamental steps: embedding representations, acquiring prototypes, and aggregating semantics. These mechanisms guarantee that MetaCluster achieves critical information extraction and reliable classification at minimal cost. The experiments encompass cybersecurity classification tasks and assess the interpretability of the framework. These tasks encompass malware family classification, threat behavior analysis, and malicious traffic identification. In particular, when compared to other DL models, MetaCluster exhibits a significant reduction in parameter consumption by 79.52% to 91.78%, and boosts operational speed up to 71.37%, while its F1 scores remain stable or slightly increase. Additionally, MetaCluster possesses the ability to assess and visually represent the significance of image, text, and statistical features. This capability leads to a reduction of Mean Squared Error (MSE) between expected and actual predictions by 0.0101 to 0.1020.
- 4 March 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Information Forensics and Security
- Vol. 19 (15566013) , 3829-3843
MetaCluster: A Universal Interpretable Classification Framework for Cybersecurity
Rising cyber threats have created an immediate demand for Deep Learning (DL) in cybersecurity. Nevertheless, the opaque nature of DL models poses challenges in deploying, collaborating, and assessing their effectiveness in less reliable cybersecurity environments. Despite eXplainable Artificial Intelligence (XAI) playing a role in enhancing cybersecurity analytics, the limited task scope, the propensity for data overfitting, and the stochastic explanations hinder its broader application. To fill the gap, this paper introduces a generic interpretable classification framework, named MetaCluster. MetaCluster generates semantic prototypes for features, patterns, and domains at varying granular levels by following three fundamental steps: embedding representations, acquiring prototypes, and aggregating semantics. These mechanisms guarantee that MetaCluster achieves critical information extraction and reliable classification at minimal cost. The experiments encompass cybersecurity classification tasks and assess the interpretability of the framework. These tasks encompass malware family classification, threat behavior analysis, and malicious traffic identification. In particular, when compared to other DL models, MetaCluster exhibits a significant reduction in parameter consumption by 79.52% to 91.78%, and boosts operational speed up to 71.37%, while its F1 scores remain stable or slightly increase. Additionally, MetaCluster possesses the ability to assess and visually represent the significance of image, text, and statistical features. This capability leads to a reduction of Mean Squared Error (MSE) between expected and actual predictions by 0.0101 to 0.1020.
- 24 October 2023
- Published by Elsevier in Computers and Electrical Engineering
- Vol. 112
STIOCS: Active learning-based semi-supervised training framework for IOC extraction
- 24 October 2023
- Published by Elsevier in Computers and Electrical Engineering
- Vol. 112
STIOCS: Active learning-based semi-supervised training framework for IOC extraction