Junjiang He
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Junjiang He has a total of 95 co-authors. Here are the top co-authors by number of shared publications:
- Xiaolong Lan
Sichuan University
16 shared publications Tao Li
School of Cyber Science and Engineering, Sichuan University, Chengdu, China
15 shared publications- Wenbo Fang
Sichuan University
15 shared publications - Jiangchuan Chen
School of Cyber Science and Engineering, Sichuan University, Chengdu, China
8 shared publications - Tao Li
Sichuan University
8 shared publications - Wenshan Li
School of Cyber Science and Engineering, Sichuan University, Chengdu, China
7 shared publications
Latest Publications
- Open Access
- 12 September 2026
- Published by Elsevier in SSRN Electronic Journal
Multimodal Malicious Website Detection via LLM-Empowered Semantic Alignment and Collaborative Reasoning
Malicious websites pose significant threats through phishing attacks, illegal content dissemination, and online fraud. Although recent studies have incorporated multimodal information and large language models (LLMs), most existing approaches rely on simple feature fusion strategies and fail to effectively capture semantic consistency across modalities. To support multimodal malicious website detection research, we first construct MW-Dataset, a multimodal malicious website dataset containing HTML source code, JavaScript/CSS files, URL information, and webpage screenshots. We further propose a Cross-Modal Semantic Alignment and Collaborative Reasoning Framework, which jointly models HTML content, JavaScript/CSS code, URLs, visual content, and OCR-extracted text through cross-modal semantic alignment and leverages LLM for collaborative reasoning and explanation generation. Experimental results on MW-Dataset and two public datasets demonstrate that the proposed method achieves superior performance among the evaluated methods, while LLM-empowered reasoning effectively improves the detection of uncertain samples near the decision boundary.
- 12 September 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessMultimodal Malicious Website Detection via LLM-Empowered Semantic Alignment and Collaborative Reasoning
Malicious websites pose significant threats through phishing attacks, illegal content dissemination, and online fraud. Although recent studies have incorporated multimodal information and large language models (LLMs), most existing approaches rely on simple feature fusion strategies and fail to effectively capture semantic consistency across modalities. To support multimodal malicious website detection research, we first construct MW-Dataset, a multimodal malicious website dataset containing HTML source code, JavaScript/CSS files, URL information, and webpage screenshots. We further propose a Cross-Modal Semantic Alignment and Collaborative Reasoning Framework, which jointly models HTML content, JavaScript/CSS code, URLs, visual content, and OCR-extracted text through cross-modal semantic alignment and leverages LLM for collaborative reasoning and explanation generation. Experimental results on MW-Dataset and two public datasets demonstrate that the proposed method achieves superior performance among the evaluated methods, while LLM-empowered reasoning effectively improves the detection of uncertain samples near the decision boundary.
- 3 September 2026
- Published by Springer Nature
- p. C1
Correction to: ProActive-Shard: A GNN and RL-Enabled Proactive Sharding Framework for Load Balancing in Blockchain
- 3 September 2026
- Published by Springer Nature
- p. C1
Correction to: ProActive-Shard: A GNN and RL-Enabled Proactive Sharding Framework for Load Balancing in Blockchain
- 21 July 2026
- Published by Springer Nature
- p. 133-144
ProActive-Shard: A GNN and RL-Enabled Proactive Sharding Framework for Load Balancing in Blockchain
- 21 July 2026
- Published by Springer Nature
- p. 133-144
ProActive-Shard: A GNN and RL-Enabled Proactive Sharding Framework for Load Balancing in Blockchain
- 11 March 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Information Forensics and Security
- Vol. 21 (15566013) , 3137-3152
Exploratory Detection of Unknown Cyber-Attacks via Evolutionary Strategy and Machine Learning
With the open-source development of cyber-attack technologies, attackers’ ability to modify existing strategies and exploit vulnerabilities has increased, leading to numerous unknown cyber-attacks. Traditional detection methods face two main challenges: (a) requiring abundant labeled attack samples, which deep learning-based detection methods find difficult to obtain in practice, and (b) struggling to effectively detect novel and previously unseen attacks, especially those that are unknown. In this paper, we propose Exploratory Detection of Unknown Cyber-Attacks via Evolutionary Strategy and Machine Learning. Specifically, firstly, we train kernel-based Ramp-OCSVM models on full features of known attacks to derive class-specific thresholds, while inferring unknown attack thresholds via Gaussian distribution. Next, we define known sample features as “genes” and generate evolutionary feature representations through multi-strategy evolution. Subsequently, these features are processed by the trained Ramp-OCSVM and the thresholds to separate known-attack variants from unknown samples. Finally, we iteratively train a RF classifier using evolved features, selecting the optimal iteration-trained model based on detection performance. We conducted extensive experiments on authoritative datasets. The results achieves F1 scores of 82.70% and 87.64% for detecting unknown attack under different configurations. The mean F1 scores improve to 99.84% and 95.80% for detecting known and unknown attacks in the few-shot learning scenario. Compared to SOTA methods, our proposed method achieves an increase of 2.19% in the F1 score, while demonstrating 53.99% higher F1-score than detection methods via GAN and VAE.
- 11 March 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Information Forensics and Security
- Vol. 21 (15566013) , 3137-3152
Exploratory Detection of Unknown Cyber-Attacks via Evolutionary Strategy and Machine Learning
With the open-source development of cyber-attack technologies, attackers’ ability to modify existing strategies and exploit vulnerabilities has increased, leading to numerous unknown cyber-attacks. Traditional detection methods face two main challenges: (a) requiring abundant labeled attack samples, which deep learning-based detection methods find difficult to obtain in practice, and (b) struggling to effectively detect novel and previously unseen attacks, especially those that are unknown. In this paper, we propose Exploratory Detection of Unknown Cyber-Attacks via Evolutionary Strategy and Machine Learning. Specifically, firstly, we train kernel-based Ramp-OCSVM models on full features of known attacks to derive class-specific thresholds, while inferring unknown attack thresholds via Gaussian distribution. Next, we define known sample features as “genes” and generate evolutionary feature representations through multi-strategy evolution. Subsequently, these features are processed by the trained Ramp-OCSVM and the thresholds to separate known-attack variants from unknown samples. Finally, we iteratively train a RF classifier using evolved features, selecting the optimal iteration-trained model based on detection performance. We conducted extensive experiments on authoritative datasets. The results achieves F1 scores of 82.70% and 87.64% for detecting unknown attack under different configurations. The mean F1 scores improve to 99.84% and 95.80% for detecting known and unknown attacks in the few-shot learning scenario. Compared to SOTA methods, our proposed method achieves an increase of 2.19% in the F1 score, while demonstrating 53.99% higher F1-score than detection methods via GAN and VAE.
- 1 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Internet of Things Journal
- Vol. 13 (17) , 1
Generating Black-Box Adversarial Examples for Industrial Control Systems via Immune Co-Evolution
- 1 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Internet of Things Journal
- Vol. 13 (17) , 1
Generating Black-Box Adversarial Examples for Industrial Control Systems via Immune Co-Evolution
- 17 December 2025
- Published by Elsevier in Applied Soft Computing
- Vol. 189
Bearing fault diagnosis with class-aware re-weighting based multi-source domain adaptation under imperfect samples
- 17 December 2025
- Published by Elsevier in Applied Soft Computing
- Vol. 189
Bearing fault diagnosis with class-aware re-weighting based multi-source domain adaptation under imperfect samples
- 17 October 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Internet of Things Journal
- Vol. 12 (24) , 55120-55136
An Immune Memory-Empowered SCADA-Based Industrial Virus Dynamic Repropagation Model
Supervisory control and data acquisition (SCADA) systems, as the core of industrial control systems and widely deployed in the nation’s critical industrial infrastructure, are attractive targets for malicious hackers due to their strategic importance. According to Check Point Research, 96% of daily cyberattacks targeting industrial control systems worldwide are known to be repeat attacks. Although current research on virus propagation assists operators in mitigating the damage caused by industrial viruses to SCADA systems, these modeling methods often fail to distinguish between initial and secondary virus invasions, making them unsuitable for modeling the repeated infection spread of industrial viruses. In order to solve this problem, we propose an immune memory-empowered SCADA-based industrial virus dynamic repropagation model, MLBRM (Memory- Latent- Broken- Robust- Memory). First, by introducing an M node, the model is used to realize the function of memorizing viral strains and to quickly immunize against and eliminate them. Besides, we perform dynamic analysis of the model and conduct the second invasion analysis to demonstrate the effect of the M nodes on suppressing the spread of the virus. In addition, we conduct a model comparison experiment and perform simulations on the U.S. power grid real dataset to demonstrate the effectiveness of the proposed model. Finally, we draw a conclusion and provide some advice for SCADA network operators to better protect the SCADA systems.
- 17 October 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Internet of Things Journal
- Vol. 12 (24) , 55120-55136
An Immune Memory-Empowered SCADA-Based Industrial Virus Dynamic Repropagation Model
Supervisory control and data acquisition (SCADA) systems, as the core of industrial control systems and widely deployed in the nation’s critical industrial infrastructure, are attractive targets for malicious hackers due to their strategic importance. According to Check Point Research, 96% of daily cyberattacks targeting industrial control systems worldwide are known to be repeat attacks. Although current research on virus propagation assists operators in mitigating the damage caused by industrial viruses to SCADA systems, these modeling methods often fail to distinguish between initial and secondary virus invasions, making them unsuitable for modeling the repeated infection spread of industrial viruses. In order to solve this problem, we propose an immune memory-empowered SCADA-based industrial virus dynamic repropagation model, MLBRM (Memory- Latent- Broken- Robust- Memory). First, by introducing an M node, the model is used to realize the function of memorizing viral strains and to quickly immunize against and eliminate them. Besides, we perform dynamic analysis of the model and conduct the second invasion analysis to demonstrate the effect of the M nodes on suppressing the spread of the virus. In addition, we conduct a model comparison experiment and perform simulations on the U.S. power grid real dataset to demonstrate the effectiveness of the proposed model. Finally, we draw a conclusion and provide some advice for SCADA network operators to better protect the SCADA systems.