DYNAMO: Towards Network Attack Campaign Attribution via Density-Aware Active Learning

Helene Orsini, Yufei HanSECRYPT 2024 · DOI

TL;DR

This paper introduces DYNAMO, a novel framework for automatically attributing network attacks to specific campaigns. It tackles the challenges of limited labeled data and imbalanced datasets by combining self-supervised learning with a density-aware active learning approach, achieving high accuracy with minimal human annotation.

What Problem Does This Paper Address?

Identifying and understanding the origins of cyberattacks (attack attribution) is crucial for effective defense. Traditional machine learning methods for this task struggle with several key issues:

  • Laborious Annotation: They require vast amounts of meticulously labeled data, which is expensive and time-consuming to obtain.
  • Imbalanced Data: Attack campaigns vary greatly in volume, leading to datasets where some campaigns are heavily overrepresented, biasing models.
  • Concept Drift: Attackers constantly evolve their tactics, making models trained on historical data quickly outdated.

This paper aims to overcome these limitations by developing a system that can accurately attribute attacks with significantly less human effort and better handle the dynamic nature of cyber threats.

Methods

DYNAMO is a three-module framework:

  1. Self-Supervised Feature Learning: It uses a self-supervised learning approach (specifically, a GraphSage-based autoencoder) to transform raw network traffic data (NetFlow records) into a lower-dimensional "latent feature space." This space is designed to group similar network traffic patterns together, making it easier to distinguish between different attack behaviors. This module reduces the reliance on labeled data by learning meaningful representations from unlabeled traffic.
  2. Density-Aware Active Learning: This module intelligently selects a small subset of the unlabeled network traffic data for human annotation. It first clusters the data in the learned feature space. Then, it prioritizes sampling from clusters that are less dense or where the current model is uncertain, ensuring that rare but important attack behaviors are not overlooked. This is a key innovation over standard active learning, which might focus too much on dense areas.
  3. Attack Attribution and Unseen Campaign Detection: DYNAMO trains two models concurrently: one to classify traffic into known attack campaigns and another to detect entirely new, previously unseen campaigns. It leverages Positive-Unlabeled (PU) learning for unseen campaign detection, treating labeled data as positive examples and a mix of known and unknown unlabeled data as potentially noisy negatives.

Datasets

The primary dataset used for evaluation is the CTU-13 dataset. This dataset contains NetFlow records from 13 different botnet scenarios, representing a large-scale, imbalanced real-world botnet traffic dataset. The authors split this dataset to simulate realistic scenarios where some campaigns are known and others are unseen.

Key Insights and Ideas

  • Feature Learning is Key: The self-supervised feature learning module significantly boosts attribution accuracy, even when using raw network traffic data. This suggests that learning a good representation of network traffic is more important than simply having more labeled data.
  • Density Matters in Active Learning: Simply picking uncertain samples (standard active learning) isn't enough. DYNAMO's density-aware approach, which prioritizes sparser regions in the feature space, leads to more balanced and representative labeled datasets, improving both attribution and unseen campaign detection.
  • Unseen Campaigns are Detectable with Minimal Labels: The framework can effectively identify new attack campaigns with very few labeled examples, demonstrating its utility in adapting to evolving threats.
  • Combining Techniques is Powerful: The synergy between self-supervised learning, density-aware active learning, and PU learning is crucial for achieving high performance across both known attack attribution and unseen campaign detection.

Key Results

  • Attack Attribution: DYNAMO, using only 3% of the network traffic records for labeling, achieved attribution accuracy comparable to fully supervised methods trained on over 80% of the data. It outperformed standard active learning (UAL) and random selection by a significant margin (up to 15% higher Macro F1 scores).
  • Unseen Campaign Detection: DYNAMO achieved near-perfect detection accuracy (Macro F1 of 1.0) for unseen campaigns with as little as 0.3% of labeled data, outperforming traditional one-class anomaly detection methods (Isolation Forest, OCSVM) by a substantial margin.
  • Data Imbalance Handling: The density-aware active learning strategy resulted in a much more balanced distribution of labeled samples across minority and majority campaigns compared to random selection.

Limitations and Future Work

  • Limitations:
    • The paper acknowledges that differentiating between very similar botnet campaigns (e.g., those using the same malware and communication patterns) can still be challenging and may require deeper payload analysis.
  • While DYNAMO handles imbalanced data well, the performance of PU learning is contingent on the representativeness of the labeled data.
  • Future Work:
    • The authors plan to extend DYNAMO's capabilities to include APT (Advanced Persistent Threat) attack campaign categorization.
  • They aim to explore self-supervised techniques for direct campaign identification, further enhancing the system's autonomy.
  • Future work will also investigate deeper analysis of communication payloads to improve fine-grained attribution.

Summary updated

BibTeX
@inproceedings{orsini2024dynamo,
  author = {Helene Orsini and Yufei Han},
  title = {{DYNAMO: Towards Network Attack Campaign Attribution via Density-Aware Active Learning}},
  booktitle = {SECRYPT},
  year = {2024},
  doi = {10.5220/0012759100003767},
}
Figure