APT Attack Detection Based on Graph Convolutional Neural Networks

Weiwu Ren, Xintong Song, Hong Yu, Lei Ying, Jinyu Yao, Yazhou Du, Wenjuan LiInternational Journal of Computational Intelligence Systems 2023 · DOI

Abstract

Advanced persistent threat (APT) attacks are malicious and targeted forms of cyberattacks that pose significant challenges to the information security of governments and enterprises. Traditional detection methods struggle to extract long-term relationships within these attacks effectively. This paper proposes an APT attack detection model based on graph convolutional neural networks (GCNs) to address this issue. The aim is to detect known attacks based on vulnerabilities and attack contexts. We extract organization-vulnerability relationships from publicly available APT threat intelligence, along with the names and relationships of software security entities from CVE, CWE, and CAPEC, to generate triple data and construct a knowledge graph of APT attack behaviors. This knowledge graph is transformed into a homogeneous graph, and GCNs are employed to process graph features, enabling effective APT attack detection. We evaluate the proposed method on the dataset constructed in this paper. The results show that the detection accuracy of the GCN method reaches 95.9%, improving by approximately 2.1% compared to the GraphSage method. This approach proves to be effective in real-world APT attack detection scenarios.

Introduction

Advanced Persistent Threats (APTs) represent a sophisticated class of cyberattacks characterized by their stealth, persistence, and specific targeting of high-value assets. Unlike common malware, APTs operate over extended periods, making them difficult to detect using traditional methods like signature-based analysis or isolated log monitoring. These conventional approaches often fail to capture the "contextual correlation" between seemingly unrelated events or vulnerabilities exploited across different stages of an attack campaign.

The primary challenge in APT detection lies in connecting fragmented pieces of threat intelligence—often found in unstructured reports—into a coherent picture of attacker behavior. This paper addresses this by leveraging Knowledge Graphs (KGs) to represent the complex relationships between vulnerabilities, weaknesses, and attack patterns, and using Graph Convolutional Networks (GCNs) to identify malicious activity within these structures.

Methodology

The proposed detection framework follows a multi-stage pipeline designed to transform raw threat intelligence into actionable detection results.

The overall architecture of the proposed APT detection framework, showing the flow from data collection to knowledge graph generation and final attack detection.
  1. Data Collection and Knowledge Graph Construction: The process begins by aggregating data from public security databases (CVE, CWE, CAPEC) and over 1,350 APT intelligence reports. From this, "triplets" (Entity-Relationship-Entity) are extracted to build a comprehensive vulnerability knowledge graph.
  2. Graph Transformation: To facilitate processing by standard GCNs, the multi-modal knowledge graph is simplified into a homogeneous graph. In this representation, nodes typically represent CVE vulnerabilities, while edges signify shared underlying weaknesses or common attack techniques.
  3. Feature Engineering: Node features are generated using a Bag-of-Words (BoW) natural language processing approach applied to the textual descriptions of vulnerabilities. This ensures that the neural network considers the semantic content of the security entities.
  4. GCN Training and Detection: The core detection engine is a GCN that performs information propagation across the graph. By aggregating features from a node's neighborhood, the GCN learns the structural context of APT attacks, allowing it to classify nodes or subgraphs as part of an ongoing threat.

Experimental Setup and Data

Due to the lack of standard benchmarks for graph-based APT detection, the researchers developed a custom dataset to evaluate the model's performance.

  • Data Sources: Intelligence was aggregated from NVD, CNNVD, Qihoo 360, FireEye, and other major global security providers.
  • Dataset Scale: The final dataset comprises 2,435 labeled attack samples and 215 distinct entity relationships derived from the analyzed reports.
  • Coverage: The data encompasses the exploitation techniques and historical behaviors of 12 prominent APT organizations.

Results and Discussion

The experimental results demonstrate the effectiveness of the GCN-based approach compared to other graph learning methods.

  • Superior Accuracy: The GCN model achieved a peak detection accuracy of 95.9%, outperforming the GraphSAGE method (93.8%). This improvement is attributed to the GCN's ability to perform convolutions over the global graph structure, whereas GraphSAGE's neighborhood sampling can lead to information loss in sparse threat intelligence graphs.
  • Performance Metrics: The model showed robust performance across various metrics, including a Precision of 95.1%, Recall of 95.8%, and an F1-score of 95.1%.
  • Efficiency: Despite its global convolution approach, the GCN model proved computationally efficient, maintaining lower average GPU utilization (9.4%) than GraphSAGE (11.4%) as the data scale increased.

Conclusion and Future Work

This research highlights the potential of combining Knowledge Graphs with Graph Convolutional Networks for identifying complex, multi-stage cyberattacks. By capturing the strategic relationships between vulnerabilities, the model provides a more holistic view of APT behaviors than traditional event-based systems.

However, limitations remain. The current simplification of heterogeneous graphs into homogeneous ones may discard valuable relational nuances. Furthermore, the model is currently static and requires retraining to adapt to zero-day threats. Future research will focus on implementing Heterogeneous Graph Neural Networks (HGNNs) to handle complex data directly and developing more adaptive mechanisms for real-time threat intelligence integration.

Summary updated

Abstract

Abstract Advanced persistent threat (APT) attacks are malicious and targeted forms of cyberattacks that pose significant challenges to the information security of governments and enterprises. Traditional detection methods struggle to extract long-term relationships within these attacks effectively. This paper proposes an APT attack detection model based on graph convolutional neural networks (GCNs) to address this issue. The aim is to detect known attacks based on vulnerabilities and attack contexts. We extract organization-vulnerability relationships from publicly available APT threat intelligence, along with the names and relationships of software security entities from CVE, CWE, and CAPEC, to generate triple data and construct a knowledge graph of APT attack behaviors. This knowledge graph is transformed into a homogeneous graph, and GCNs are employed to process graph features, enabling effective APT attack detection. We evaluate the proposed method on the dataset constructed in this paper. The results show that the detection accuracy of the GCN method reaches 95.9%, improving by approximately 2.1% compared to the GraphSage method. This approach proves to be effective in real-world APT attack detection scenarios.
BibTeX
@article{ren2023apt,
  author = {Weiwu Ren and Xintong Song and Hong Yu and Lei Ying and Jinyu Yao and Yazhou Du and Wenjuan Li},
  title = {{APT Attack Detection Based on Graph Convolutional Neural Networks}},
  journal = {International Journal of Computational Intelligence Systems},
  year = {2023},
  doi = {10.1007/s44196-023-00369-5},
}
Figure