THREATRACE: Detecting and Tracing Host-Based Threats in Node Level Through Provenance Graph Learning

Su Wang, Zhiliang Wang, Tao Zhou, Hongbin Sun, Xia Yin, Dongqi Han, Han Zhang, Xingang Shi, Jiahai Yang

TL;DR

This paper introduces THREATRACE, a novel system for detecting and tracing stealthy host-based threats by learning the "normal" behavior of individual system entities (nodes) within a data provenance graph. By leveraging GraphSAGE, THREATRACE can identify subtle anomalies that traditional graph-level or misuse-based methods miss, offering improved detection and tracing capabilities.

What Problem Does This Paper Address?

Modern attackers, particularly those employing Advanced Persistent Threats (APTs), are increasingly using stealthy and persistent tactics to infiltrate systems. These attacks often exploit zero-day vulnerabilities, making signature-based (misuse) detection ineffective. While anomaly-based methods using data provenance (a graph representing system activities) have shown promise, existing approaches often focus on the entire provenance graph. This is problematic because stealthy threats might only involve a small number of anomalous entities, making the overall graph appear similar to benign activity. Consequently, these methods struggle to detect subtle threats and, crucially, cannot pinpoint the location of the anomalous behavior, hindering effective response and remediation.

Methods

THREATRACE tackles this challenge by shifting the focus from the entire graph to individual nodes (system entities like processes or files). The core of THREATRACE is a tailored GraphSAGE-based framework, a type of Graph Neural Network (GNN) capable of learning node representations within a graph.

Key methodological innovations include:

  • Node-Level Anomaly Detection: Instead of classifying entire graphs, THREATRACE learns the "normal" role and behavior of each benign node. Anomalies are detected when a node's behavior deviates significantly from its learned benign profile.
  • Inductive Learning with GraphSAGE: GraphSAGE is chosen for its ability to learn an embedding function that can generalize to new, unseen nodes and evolving graphs, making it suitable for real-time, long-term monitoring.
  • Multi-Model Framework: To address the data imbalance problem (where anomalous nodes are rare) and the diversity of benign node behaviors, THREATRACE employs a multi-model approach. It trains multiple specialized submodels, each learning the characteristics of different types of benign nodes. This allows for more nuanced detection, especially for nodes with unique or less frequent benign roles.
  • Probability-Based Detection: A probability-based method is used during both training and execution to refine anomaly detection. This helps reduce false positives by requiring a high degree of confidence for a node to be classified as benign.
  • Anomaly Tracing: When an anomaly is detected, THREATRACE can directly trace its location within the provenance graph by examining the node's local neighborhood (2-hop ancestors and descendants).

Datasets

The paper evaluates THREATRACE on three public datasets:

  • StreamSpot Dataset: Used for comparing with StreamSpot and Unicorn.
  • Unicorn SC-2 Dataset: A more complex dataset used for comparing with Unicorn and ProvDetector.
  • DARPA TC Dataset: A real-world dataset generated from a red-team vs. blue-team engagement, providing ground truth for evaluating node-level detection and tracing capabilities.

Key Insights and Ideas

  • Focusing on the Node, Not the Graph: The most significant insight is that even stealthy attacks leave subtle deviations in the behavior of individual system entities. By learning these "normal" node profiles, THREATRACE can detect anomalies that are invisible at the graph level.
  • Multi-Model Learning for Imbalance: The multi-model approach is crucial for handling the inherent data imbalance in security datasets. It allows the system to learn the nuances of various benign node types, preventing rare but legitimate behaviors from being flagged as malicious.
  • Traceability is Key: The ability to not only detect but also trace the anomalous behavior to specific nodes is vital for security analysts. THREATRACE's node-level approach directly enables this crucial functionality.
  • GraphSAGE's Suitability: The inductive nature of GraphSAGE makes it well-suited for dynamic, streaming provenance data, where the graph structure can evolve over time.

Key Results

  • Superior Detection Performance: THREATRACE consistently outperformed state-of-the-art detectors like Unicorn and ProvDetector on the StreamSpot and Unicorn SC-2 datasets, achieving high precision and recall.
  • Effective Anomaly Tracing: On the DARPA TC dataset, THREATRACE demonstrated strong performance in detecting and tracing anomalous nodes, achieving high F-scores and low false positive rates.
  • Robustness: The system showed good robustness against optimization-based evasion attacks, indicating its resilience to adversaries attempting to mask malicious activity.
  • Efficiency: THREATRACE offers a good balance between detection performance and runtime overhead, making it suitable for real-time deployment.

Limitations and Future Work

  • Closed-World Assumption: Like most anomaly detection systems, THREATRACE operates under a closed-world assumption, meaning it assumes all benign behaviors are covered in the training data. Real-world systems may present novel benign behaviors not seen during training. The authors suggest periodic model updates with new benign data.
  • Granularity of Data Provenance: Certain threats, like malicious code within files or thread-level activities, might not leave significant traces in the current granularity of data provenance. Future work could involve improving provenance granularity or integrating other data sources.
  • System Overhead: While efficient, THREATRACE, being a deep-learning-based method, has higher computation resource requirements than simpler methods. Adjusting parameters like Batch Size (BS) and Subgraph Size (SS) can help balance performance and overhead.
  • False Positives: Despite efforts to reduce them, THREATRACE can still generate false positives, particularly on datasets with a large number of benign nodes. Future work could explore whitelist mechanisms or adaptive training to further mitigate this.
  • Attack Story Construction: While THREATRACE can trace anomalies, it doesn't construct a full "attack story." This is identified as a gap for future research.

Summary updated

BibTeX
@article{wang2022threatrace,
  author = {Su Wang and Zhiliang Wang and Tao Zhou and Hongbin Sun and Xia Yin and Dongqi Han and Han Zhang and Xingang Shi and Jiahai Yang},
  title = {{THREATRACE: Detecting and Tracing Host-Based Threats in Node Level Through Provenance Graph Learning}},
  journal = {IEEE Trans. Inf. Forensics Secur.},
  year = {2022},
  url = {https://openreview.net/forum?id=ZO8rgiM7IF},
}
Figure