An Improved NPCUSUM Method with Adaptive Sliding Window to Detect DDoS Attacks

Degang Sun, Kun Yang, Weiqing Huang, Yan Wang, Bo HuLecture Notes in Computer Science 2016 · DOI · OpenAlex

Abstract

Detecting Distributed Denial of Service (DDoS) attacks in real-time requires a balance between computational efficiency and sensitivity to dynamic network conditions. This paper proposes an enhanced detection framework that integrates Non-parametric Cumulative Sum (NPCUSUM) statistics with an Adaptive Sliding Window (ASW) mechanism. By utilizing Netflow data instead of raw packet captures, the system achieves significant speed improvements. The inclusion of a "tolerance factor" and automated threshold updates allows the model to adapt to fluctuating traffic patterns, reducing false positives while maintaining high detection accuracy.

Introduction

DDoS attacks continue to pose a significant threat to global internet infrastructure. Traditional detection methods often struggle with three primary issues: high computational overhead from analyzing raw packet data, rigid performance due to static window sizes and thresholds, and a reliance on manual parameter tuning. As network traffic is inherently volatile, static systems frequently suffer from poor scalability and high false-alarm rates. This research addresses these gaps by developing a flexible, automated, and efficient detection pipeline.

Methodology

The proposed system leverages a multi-stage approach to transform raw network data into actionable security alerts.

Data Processing and Feature Selection

To ensure real-time performance, the system converts raw pcap data into Netflow records. This abstraction significantly reduces data volume while preserving critical features. The model focuses on two key metrics: Conditional Entropy ($H(srcIP|dstIP)$), which measures the uncertainty of source IP addresses targeting a specific destination, and Flow Count ($flowCnt$).

Adaptive Sliding Window and NPCUSUM

The core of the detection engine is the Adaptive Sliding Window (ASW). Unlike fixed-size windows, the ASW dynamically adjusts its size based on the current traffic state. It expands during periods of normal traffic to establish a stable baseline and contracts when an anomaly is suspected to isolate the potential attack.

Architecture of the proposed NPCUSUM method with Adaptive Sliding Window (ASW).

The system applies an improved NPCUSUM algorithm to these windows, calculating the cumulative sum of deviations from the learned "normal" mean. To mitigate "jitter" and transient spikes, a tolerance factor is employed: an alert is only triggered if anomalies persist beyond a specific threshold within a set timeframe.

Key Insights and Results

The researchers evaluated the framework using the DARPA 2000 Intrusion Detection Scenario Specific Data Set.

Efficiency and Speed

The transition to Netflow-based analysis yielded dramatic performance gains. The system processed a three-hour traffic capture in approximately 33 minutes (1993.35 seconds), outperforming traditional entropy-based methods that often require several hours for the same task.

Dynamic Adaptability

A critical finding is the system's ability to maintain a "moving target" baseline. By allowing thresholds to evolve with the traffic, the model avoids the obsolescence typical of static models.

Dynamic threshold adjustments for Source-Destination IP Entropy and Flow Count.

As shown in the experimental results, the thresholds for both entropy and flow count successfully adjusted during peak traffic periods. This adaptability, combined with the tolerance factor, ensures that the system remains robust against legitimate network fluctuations while accurately identifying sustained malicious activity.

Conclusion

The improved NPCUSUM method with an adaptive sliding window offers a scalable solution for real-time DDoS detection. While highly effective, the authors note that the system currently faces challenges in distinguishing between malicious attacks and "Flash Crowds" (legitimate traffic surges). Additionally, the tolerance factor introduces a minor detection delay. Future research will focus on refining the algorithm to differentiate between these scenarios and further optimizing alert latency.

Summary updated

BibTeX
@inproceedings{sun2016improved,
  author = {Degang Sun and Kun Yang and Weiqing Huang and Yan Wang and Bo Hu},
  title = {{An Improved NPCUSUM Method with Adaptive Sliding Window to Detect DDoS Attacks}},
  booktitle = {Lecture Notes in Computer Science},
  year = {2016},
  doi = {10.1007/978-3-319-29814-6\_25},
}
Figure