Network security situation prediction based on singular spectrum analysis

Jizhi Wang, Bo Hu, Jianguo Jiang, Jiqiang Liu2016 · DOI

Abstract

Network security situation prediction based on historical data is intended to provide advance warnings to network administrators. However, current approaches often ignore the fact that original security time series include random components that are inherently unforecastable. This paper utilizes Singular Spectrum Analysis (SSA) to separate these random components. Subsequently, an ARMA model is used to predict the security situation in the near future. Experiments demonstrate that prediction accuracy benefits significantly from this separation.

Introduction

Network administrators rely on "situation awareness" to stay ahead of cyber threats. Predicting spikes in security risks allows for proactive defensive measures. However, raw network security data is a complex time series containing both predictable trends and sudden, random events—such as short-term Denial of Service (DoS) attacks. Standard forecasting models often fail because they attempt to learn from these unpredictable spikes, leading to "poisoned" historical data and wildly inaccurate future predictions. This paper addresses the challenge of separating meaningful signals from random noise to improve prediction reliability.

Methodology: SSA and ARMA

The authors propose a two-stage pipeline to enhance prediction accuracy. The core innovation is the use of Singular Spectrum Analysis (SSA) as a pre-processing filter. SSA is a non-parametric technique that decomposes a time series into a predictable trend component and a random noise component.

SSA decomposition: The top plot shows the raw data (blue) vs. the extracted trend (red); the bottom plot shows the isolated random noise (green).

The SSA process involves four primary steps: Embedding (creating a trajectory matrix), Singular Value Decomposition (SVD), Grouping (separating signal from noise), and Reconstruction. Once the noise is stripped away, a standard Auto-Regressive Moving Average (ARMA) model is employed to forecast future values based solely on the cleaned trend component. This ensures the model is not "confused" by transient anomalies.

Experimental Results

The study utilized a real-world dataset from an organization's network, consisting of 48 data points collected every 30 minutes over a 24-hour period. A significant DoS attack occurred at the 43rd data point, creating a sharp spike in the normalized security situation value.

The researchers compared three prediction scenarios:

  • P1 (Baseline): Prediction using data prior to the DoS spike, yielding reasonable accuracy for the general trend.
  • P2 (Raw Data): Prediction using data that included the DoS spike without filtering. The ARMA model interpreted the spike as a new trend, resulting in completely erroneous forecasts.
  • P3 (SSA-Filtered): Prediction using the same data as P2, but after SSA filtering.
The P3 prediction (red) successfully ignores the random spike at time 43, maintaining a stable and accurate forecast of the underlying security trend.

As shown in the results, SSA successfully isolated the DoS spike into the noise component. Consequently, the P3 prediction remained stable and closely aligned with the actual post-attack trend, demonstrating that predictability is significantly improved by acknowledging and removing "no-sense" random data.

Conclusion

By integrating SSA with ARMA, this approach provides a robust framework for network security forecasting. While the current study is limited by a small dataset and offline analysis, it highlights the critical importance of noise separation. Future work should focus on scaling this method for high-frequency, real-time monitoring and optimizing parameter selection for dynamic network environments.

Summary updated

Abstract

Network security situation prediction based on historical data is to provide warning in advance with network administrator. However, the current proposed approaches ignores the fact that the original time series of security situation include the random component which is not able to be forecast. Singular spectrum analysis is utilized to separate the random component. After that, ARMA model is used to predicate the security situation value in near future. The experiment shows the prediction benefits from the separation.
BibTeX
@misc{wang2016network,
  author = {Jizhi Wang and Bo Hu and Jianguo Jiang and Jiqiang Liu},
  title = {{Network security situation prediction based on singular spectrum analysis}},
  year = {2016},
  doi = {10.2991/wartia-16.2016.352},
}
Figure