Kai Shi
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
Kai Shi has a total of 50 co-authors. Here are the top co-authors by number of shared publications:
- Jinsong Wang
School of Computer Science and Engineering, Tianjin University of Technology, Tianjin, China
5 shared publications Jinsong Wang
Tianjin University of Technology
4 shared publicationsAnxing Jiang
Tianjin University of Technology,School of Computer Science and Engineering,Tianjin,China
3 shared publications- Mengshu Wang
Tianjin University of Technology,School of Computer Science and Engineering,Tianjin,China
3 shared publications Oliver Yang
School of Information Technology and Engineering, University of Ottawa, Ottawa, ON, Canada
3 shared publicationsSheng Lin
Tianjin University of Technology
3 shared publications
Latest Publications
- 6 November 2026
- Published by SAGE Publications in Intelligent Data Analysis: An International Journal
EFGTC: An explainable federated graph learning framework for encrypted traffic classification
With the development of Decentralized Applications (DApps) and Virtual Private Networks (VPNs), encrypted traffic classification has become a core task of network security supervision. However, the implementation of data privacy regulations restricts the application of traditional centralized classification methods, making the problem of data silos increasingly prominent. Although Federated Learning (FL) provides a privacy-preserving paradigm for solving this problem, existing FL-based schemes mostly ignore the intrinsic topological structure of traffic data, leading to difficulties in extracting comprehensive features and a lack of interpretability in model decisions. To address these issues, this paper proposes EFGTC , an Explainable Federated Graph Learning Framework for Encrypted Traffic Classification. First, we propose a time-aware weighted graph construction method. Unlike traditional methods that purely use connection relationships, we construct a weighted topology graph rich in temporal rhythm by jointly encoding packet direction, length, and inter-arrival time, thereby enabling the graph data to more accurately express the interaction patterns of encrypted traffic. Second, we design a multi-scale feature extraction module based on an attention mechanism. This module not only captures comprehensive features from local handshake details to global flow patterns through multi-level aggregation but also achieves automatic screening and focusing of key data packets through an attention pooling mechanism, effectively suppressing data noise. Finally, experiments on three public datasets within the FL framework show that EFGTC consistently outperforms baselines across all metrics and exhibits strong robustness. Attention-weight visualizations further confirm that the model identifies key traffic-category fingerprints, enhancing interpretability.
- 6 November 2026
- Published by SAGE Publications in Intelligent Data Analysis: An International Journal
EFGTC: An explainable federated graph learning framework for encrypted traffic classification
With the development of Decentralized Applications (DApps) and Virtual Private Networks (VPNs), encrypted traffic classification has become a core task of network security supervision. However, the implementation of data privacy regulations restricts the application of traditional centralized classification methods, making the problem of data silos increasingly prominent. Although Federated Learning (FL) provides a privacy-preserving paradigm for solving this problem, existing FL-based schemes mostly ignore the intrinsic topological structure of traffic data, leading to difficulties in extracting comprehensive features and a lack of interpretability in model decisions. To address these issues, this paper proposes EFGTC , an Explainable Federated Graph Learning Framework for Encrypted Traffic Classification. First, we propose a time-aware weighted graph construction method. Unlike traditional methods that purely use connection relationships, we construct a weighted topology graph rich in temporal rhythm by jointly encoding packet direction, length, and inter-arrival time, thereby enabling the graph data to more accurately express the interaction patterns of encrypted traffic. Second, we design a multi-scale feature extraction module based on an attention mechanism. This module not only captures comprehensive features from local handshake details to global flow patterns through multi-level aggregation but also achieves automatic screening and focusing of key data packets through an attention pooling mechanism, effectively suppressing data noise. Finally, experiments on three public datasets within the FL framework show that EFGTC consistently outperforms baselines across all metrics and exhibits strong robustness. Attention-weight visualizations further confirm that the model identifies key traffic-category fingerprints, enhancing interpretability.
- 3 February 2026
- Published by Springer Nature in Multimedia Systems
- Vol. 32 (2) , 88
BDFL: a blockchain-based decentralized federated learning security framework
- 3 February 2026
- Published by Springer Nature in Multimedia Systems
- Vol. 32 (2) , 88
BDFL: a blockchain-based decentralized federated learning security framework
- 31 December 2025
- Published by Elsevier in Ad Hoc Networks
- Vol. 183
UpsFed-IDS: U-shaped split federated intrusion detection system for securing UAV communication in dynamic networks
- 31 December 2025
- Published by Elsevier in Ad Hoc Networks
- Vol. 183
UpsFed-IDS: U-shaped split federated intrusion detection system for securing UAV communication in dynamic networks
- 21 June 2025
- Published by Springer Nature in The Journal of Supercomputing
- Vol. 81 (9) , 1-48
A few-shot detection method for new types of network traffic attacks based on meta-learning with cross-attention
- 21 June 2025
- Published by Springer Nature in The Journal of Supercomputing
- Vol. 81 (9) , 1-48
A few-shot detection method for new types of network traffic attacks based on meta-learning with cross-attention
- 27 February 2024
- Published by Springer Nature
- p. 42-62
SolGPT: A GPT-Based Static Vulnerability Detection Model for Enhancing Smart Contract Security
- 27 February 2024
- Published by Springer Nature
- p. 42-62
SolGPT: A GPT-Based Static Vulnerability Detection Model for Enhancing Smart Contract Security
- Open Access
- 22 December 2023
- Published by Association for Computing Machinery (ACM)
- p. 1-5
A 3D CNN Model With 3d CBAM Layer For Micro-Expression Recognition
Facial microexpression recognition in videos is a popular area of computer vision research. Microexpressions are non-verbal expressions of human emotions that often appear when people try to hide or control their emotions.There are many facial expression recognition methods in related fields, but micro-expression recognition in videos has not been accurate enough. This paper aims to study micro-expression recognition technology, which can effectively identify small and short-lived expression changes on human faces. This study proposes a micro-expression recognition method based on computer vision and machine learning by analyzing facial features and spatio-temporal feature extraction. First, we collected a video dataset containing micro-expressions and performed preprocessing and feature extraction on them. Then, we use more advanced methods in the field and methods proposed by the team to train and test the model to recognize multiple micro-expression categories. Experimental results show that our method achieves higher accuracy and more stable performance in micro-expression recognition tasks.
- 22 December 2023
- Published by Association for Computing Machinery (ACM)
- p. 1-5
Open AccessA 3D CNN Model With 3d CBAM Layer For Micro-Expression Recognition
Facial microexpression recognition in videos is a popular area of computer vision research. Microexpressions are non-verbal expressions of human emotions that often appear when people try to hide or control their emotions.There are many facial expression recognition methods in related fields, but micro-expression recognition in videos has not been accurate enough. This paper aims to study micro-expression recognition technology, which can effectively identify small and short-lived expression changes on human faces. This study proposes a micro-expression recognition method based on computer vision and machine learning by analyzing facial features and spatio-temporal feature extraction. First, we collected a video dataset containing micro-expressions and performed preprocessing and feature extraction on them. Then, we use more advanced methods in the field and methods proposed by the team to train and test the model to recognize multiple micro-expression categories. Experimental results show that our method achieves higher accuracy and more stable performance in micro-expression recognition tasks.
- 25 August 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 302-306
Malicious Process Identification via System Audit Logs Analysis
Constructing a dependency graph through the utilization of system audit logs and subsequently performing causal relationship analysis on this graph has become one of the crucial methods for conducting attack investigations. The introduction of the dependency graph makes it possible to identification malicious system entities by providing correlated information within the context of the attack. In this paper, a method is proposed that combines biased random walks with deep anomaly detection models to identify malicious process entities within dependency graphs. The process begins with the utilization of system audit logs to construct a dependency graph, which is subsequently simplified. In this graph, individual edges are endowed with weights, thus forming a weighted dependency graph. Progressing further, a method of biased random walks is implemented to analyze process entities present within the dependency graph. This analysis effectively captures vital contextual information that represents process behaviors, and subsequently transforming them into embedded vectors. Finally, a deep anomaly detection model is employed to achieve the discovery of attacking processes within the system. The effectiveness of the proposed approach is evaluated on six instances of attacks using the DARPA TC dataset. The experimental results demonstrate the method's proficiency in detecting malicious system entities.
- 25 August 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 302-306
Malicious Process Identification via System Audit Logs Analysis
Constructing a dependency graph through the utilization of system audit logs and subsequently performing causal relationship analysis on this graph has become one of the crucial methods for conducting attack investigations. The introduction of the dependency graph makes it possible to identification malicious system entities by providing correlated information within the context of the attack. In this paper, a method is proposed that combines biased random walks with deep anomaly detection models to identify malicious process entities within dependency graphs. The process begins with the utilization of system audit logs to construct a dependency graph, which is subsequently simplified. In this graph, individual edges are endowed with weights, thus forming a weighted dependency graph. Progressing further, a method of biased random walks is implemented to analyze process entities present within the dependency graph. This analysis effectively captures vital contextual information that represents process behaviors, and subsequently transforming them into embedded vectors. Finally, a deep anomaly detection model is employed to achieve the discovery of attacking processes within the system. The effectiveness of the proposed approach is evaluated on six instances of attacks using the DARPA TC dataset. The experimental results demonstrate the method's proficiency in detecting malicious system entities.
- 25 August 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 429-433
Research on Network Traffic Feature Map Generation and Evaluation Methods
With the continuous development of machine learning, it has been widely used in the fields of image processing and computer vision. However, it needs to be transformed into images before feeding network traffic data into a deep learning model for training. The existing methods for directly converting network traffic data into images demonstrate unfavorable results. Therefore, this paper proposes a method for generating network traffic feature maps. This method first uses information entropy and classical k-nearest neighbors estimation algorithm to calculate the entropy values for discrete and continuous variables respectively, then the entropy values are used to filter network traffic. Finally, feature maps are generated through an encoder. Experiments on the UNSW_NB15 and CIC-IDS2017 datasets show that the proposed method outperforms other methods in terms of metrics such as PSNR and SSIM.
- 25 August 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 429-433
Research on Network Traffic Feature Map Generation and Evaluation Methods
With the continuous development of machine learning, it has been widely used in the fields of image processing and computer vision. However, it needs to be transformed into images before feeding network traffic data into a deep learning model for training. The existing methods for directly converting network traffic data into images demonstrate unfavorable results. Therefore, this paper proposes a method for generating network traffic feature maps. This method first uses information entropy and classical k-nearest neighbors estimation algorithm to calculate the entropy values for discrete and continuous variables respectively, then the entropy values are used to filter network traffic. Finally, feature maps are generated through an encoder. Experiments on the UNSW_NB15 and CIC-IDS2017 datasets show that the proposed method outperforms other methods in terms of metrics such as PSNR and SSIM.
- 25 August 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 424-428
Constructing Weighted Host Communication Networks for Intrusion Detection based on Complex Network Theory
As the scope of the Internet continues to broaden, there is an exponential surge in network data, accompanied by a simultaneous escalation of cybersecurity issues. To accurately identify the source of intrusion within a compromised network, this paper proposes an intrusion detection method based on complex network theory. In this method, the network hosts are regarded as nodes, the communication relationships between hosts are seen as edges, and the communication traffic features are mapped as weights to construct weighted host communication networks. Subsequently, complex network metrics that assess the significance of host nodes are extracted. These metrics are then combined with an anomaly detection model to achieve intrusion detection. The proposed method has been verified through testing on two datasets, namely the CIC-IDS-2017 and DAPT2020 datasets. And the experimental results substantiate that the method demonstrate a positive intrusion detection performance, outperform the compared anomaly detection models.
- 25 August 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 424-428
Constructing Weighted Host Communication Networks for Intrusion Detection based on Complex Network Theory
As the scope of the Internet continues to broaden, there is an exponential surge in network data, accompanied by a simultaneous escalation of cybersecurity issues. To accurately identify the source of intrusion within a compromised network, this paper proposes an intrusion detection method based on complex network theory. In this method, the network hosts are regarded as nodes, the communication relationships between hosts are seen as edges, and the communication traffic features are mapped as weights to construct weighted host communication networks. Subsequently, complex network metrics that assess the significance of host nodes are extracted. These metrics are then combined with an anomaly detection model to achieve intrusion detection. The proposed method has been verified through testing on two datasets, namely the CIC-IDS-2017 and DAPT2020 datasets. And the experimental results substantiate that the method demonstrate a positive intrusion detection performance, outperform the compared anomaly detection models.