GBDT-IL: Incremental Learning of Gradient Boosting Decision Trees to Detect Botnets in Internet of Things

Ruidong Chen, Tianci Dai, Yanfeng Zhang, Yukun Zhu, Xin Liu, Erfan ZhaoSensors 2024 · DOI

TL;DR

This paper introduces GBDT-IL, a novel incremental learning approach for detecting IoT botnets that effectively adapts to evolving threats and dynamic network data. By combining an enhanced Fisher Score for efficient feature selection with an incremental Gradient Boosting Decision Tree (GBDT) model, GBDT-IL achieves high accuracy with fewer features and robust performance against concept drift.

What Problem Does This Paper Address?

The Internet of Things (IoT) is rapidly expanding, bringing convenience but also significant security risks, particularly from the proliferation of botnets. While AI-based methods have shown promise in detecting these threats, they often struggle with the dynamic and evolving nature of IoT network data. This "concept drift" causes existing models to degrade in performance over time as botnets change their behavior. Furthermore, resource-constrained IoT devices pose a challenge for computationally intensive detection models. This paper addresses the need for an accurate, efficient, and robust IoT botnet detection system that can adapt to evolving threats and operate within limited resource environments.

Methods

The proposed solution, GBDT-IL, has two main components:

  1. Enhanced Fisher Score Feature Selection: The authors introduce an improved Fisher Score method to select the most informative features. This method addresses limitations of the traditional Fisher Score by better handling feature overlap and inter-class variance, aiming to reduce dimensionality while preserving discriminative power. This is crucial for resource-limited IoT devices.
  2. Incremental Gradient Boosting Decision Tree (GBDT-IL): This core component uses an incremental learning approach to adapt the GBDT model to concept drift. Instead of retraining the entire model, GBDT-IL incrementally adds new decision trees based on residuals from new data. To prevent overfitting and manage model complexity, it incorporates a pruning mechanism to remove redundant trees. This allows the model to continuously learn and adapt to new patterns in the data stream.

Datasets

The study utilized four publicly available IoT datasets for evaluation:

  • N-BaIoT: From IP camera video surveillance and IoT devices, featuring various attack types.
  • BoT-IoT: Simulated IoT environment with Raspberry Pi sensors and UNSW-NB15 attack data.
  • MedBIoT: Collected data from three botnet malware types, covering real and virtual IoT devices.
  • MQTTSet: Generated using IoT-Flock to simulate benign and attack traffic focused on the MQTT protocol.

To specifically test concept drift handling, the authors artificially constructed drift datasets based on these original datasets, simulating scenarios with changing attack patterns and distributions.

Key Insights and Ideas

  • Feature Selection is Crucial for IoT: The paper strongly emphasizes that reducing the number of features is not just about efficiency but is a necessity for deploying effective botnet detection on resource-constrained IoT devices.
  • Incremental Learning for Dynamic Threats: The core idea of GBDT-IL is that botnets are not static. An incremental learning approach is essential to allow detection models to adapt and learn new attack behaviors as they emerge, rather than becoming obsolete.
  • Pruning for Robustness: Simply adding new trees in an incremental fashion can lead to overfitting. The inclusion of a pruning step in GBDT-IL is a key insight for maintaining model performance and preventing it from becoming overly specialized to past data.
  • Fisher Score Enhancement: The authors demonstrate that even well-established methods like Fisher Score can be improved to better suit the nuances of real-world, potentially overlapping, data distributions.

Key Results

  • High Accuracy with Few Features: The enhanced Fisher Score feature selection method achieved high accuracy (e.g., 99.81% on average across datasets) using a significantly reduced feature set (e.g., 25 features). This is a substantial improvement over using all available features, leading to reduced training time and resource usage.
  • Robust Concept Drift Detection: GBDT-IL demonstrated superior performance in concept drift scenarios, achieving an average accuracy of 96.88%, outperforming existing concept drift detection algorithms by 2.98%.
  • Low False Positive Rate: The proposed method maintained a low average false positive rate of 3.02%, indicating its ability to correctly distinguish benign traffic from malicious traffic.
  • Comparative Performance: GBDT-IL consistently outperformed traditional machine learning algorithms and other concept drift-resistant algorithms across various experimental setups.

Limitations and Future Work

  • Feature Combination Optimization: The authors acknowledge that their current feature selection method prioritizes individual features based on score. They suggest future work could explore feature combination optimization to potentially yield even better results, as the best combination might not always be the top-scoring individual features.
  • Supervised Learning in Real-World Scenarios: GBDT-IL is a supervised learning algorithm. The paper notes that in real-world production, traffic might not be easily labeled as benign or malicious, posing a challenge for applying supervised methods. Future research could investigate how to effectively integrate this approach into realistic, potentially semi-supervised or unsupervised, environments.
  • Further Exploration of WFS: The paper mentions that the Weighted Feature Selection (WFS) method showed results close to their proposed NFS, particularly on image domains. They suggest that exploring how to better integrate WFS for traffic feature selection is a worthwhile avenue for investigation.

Summary updated

Abstract

The rapid development of the Internet of Things (IoT) has brought many conveniences to our daily life. However, it has also introduced various security risks that need to be addressed. The proliferation of IoT botnets is one of these risks. Most of researchers have had some success in IoT botnet detection using artificial intelligence (AI). However, they have not considered the impact of dynamic network data streams on the models in real-world environments. Over time, existing detection models struggle to cope with evolving botnets. To address this challenge, we propose an incremental learning approach based on Gradient Boosting Decision Trees (GBDT), called GBDT-IL, for detecting botnet traffic in IoT environments. It improves the robustness of the framework by adapting to dynamic IoT data using incremental learning. Additionally, it incorporates an enhanced Fisher Score feature selection algorithm, which enables the model to achieve a high accuracy even with a smaller set of optimal features, thereby reducing the system resources required for model training. To evaluate the effectiveness of our approach, we conducted experiments on the BoT-IoT, N-BaIoT, MedBIoT, and MQTTSet datasets. We compared our method with similar feature selection algorithms and existing concept drift detection algorithms. The experimental results demonstrated that our method achieved an average accuracy of 99.81% using only 25 features, outperforming similar feature selection algorithms. Furthermore, our method achieved an average accuracy of 96.88% in the presence of different types of drifting data, which is 2.98% higher than the best available concept drift detection algorithms, while maintaining a low average false positive rate of 3.02%.
BibTeX
@inproceedings{chen2024gbdt,
  author = {Ruidong Chen and Tianci Dai and Yanfeng Zhang and Yukun Zhu and Xin Liu and Erfan Zhao},
  title = {{GBDT-IL: Incremental Learning of Gradient Boosting Decision Trees to Detect Botnets in Internet of Things}},
  booktitle = {Sensors},
  year = {2024},
  doi = {10.3390/s24072083},
}
Figure