TL;DR
This paper introduces a novel framework for detecting low-intensity flooding attacks in backbone networks by combining the "Sketch" data structure with "Power Divergence" measures. This approach offers improved accuracy and reduced false alarms compared to existing methods, particularly for subtle attacks that can be masked by normal network traffic.
What Problem Does This Paper Address?
Detecting flooding attacks, especially low-intensity ones, in high-speed backbone networks is a significant challenge. These attacks aim to overwhelm network resources, denying service to legitimate users. Traditional methods often struggle because:
- Data Volume: Backbone traffic generates massive amounts of data, making per-flow analysis computationally infeasible.
- Low-Intensity Attacks: Subtle attacks can be easily masked by the natural variations and high volume of normal network traffic, leading to missed detections.
- False Alarms: Methods that aggregate traffic into a single time series are prone to false alarms due to normal traffic fluctuations, especially when using static thresholds.
The paper aims to address these issues by developing a more granular and sensitive detection mechanism that can identify these hard-to-detect attacks without generating an excessive number of false positives.
Methods
The proposed approach combines two key techniques:
- Sketch Data Structure: This technique provides a way to summarize network traffic in a fixed memory footprint, allowing for more granular analysis than simply aggregating all traffic into one time series. It uses multiple hash functions to randomly aggregate traffic (specifically, SYN requests in this paper, identified by Destination IP) into buckets, creating a compressed representation of traffic patterns. This allows for a more detailed view of traffic distributions.
- Power Divergence (PD): This is a generalized measure of divergence between probability distributions. The authors use PD to compare the probability distribution of network traffic in a current time interval with that of a previous, "normal" interval. Different values of the parameter $\beta$ in the PD formula allow it to encompass other well-known divergence measures like Kullback-Leibler (KL) divergence, Hellinger Distance (HD), and Chi-square ($\chi^2$) divergence. The core idea is that a significant deviation between these distributions, as measured by PD, indicates an anomaly or attack.
The framework works by:
- Using Sketch to create probability distributions from traffic data.
- Calculating the Power Divergence between current and historical distributions.
- Employing a dynamically updated threshold (based on EWMA) to detect deviations.
- Requiring the deviation to persist for a certain number of intervals ($\eta$) to raise an alarm, further reducing false positives.
Datasets
The paper evaluates its approach using real Internet traffic traces from the MAWI trans-Pacific wide transit link between the USA and Japan. These traces were collected from April 15, 2010, between 12:00 and 18:15. The authors also mention injecting synthetic DDoS attacks into these traces for their experiments.
Key Insights and Ideas
- Granularity is Key: Aggregating all traffic into a single time series is insufficient for detecting low-intensity attacks. The Sketch data structure's ability to provide a more granular view of traffic distributions is crucial.
- Power Divergence as a Flexible Tool: Power Divergence, with its tunable parameter $\beta$, offers a more flexible and potentially more powerful way to measure distribution differences than fixed measures like KL or HD.
- Optimizing for Detection: By varying the $\beta$ parameter, the authors demonstrate that Power Divergence can be tuned to achieve better detection rates and lower false alarm rates, outperforming its special cases.
- Dynamic Thresholding Reduces False Alarms: Relying on a static threshold is problematic. The use of a dynamically updated threshold, combined with a persistence requirement for alarms, significantly helps in distinguishing real attacks from normal traffic variations.
Key Results
- The proposed Power Divergence over Sketch framework successfully detects low-intensity SYN flooding attacks.
- The authors found that $\beta = 2.5$ for Power Divergence provided the optimal performance, achieving a 100% True Positive Rate (DR) with a 10% False Alarm Rate (FAR).
- For comparison, other values of $\beta$ yielded:
- $\beta = 0.5$ (proportional to HD): 100% DR, 30% FAR
- $\beta = 1$ (KL divergence): 100% DR, 23% FAR
- $\beta = 1.5$ and $\beta = 2$ (proportional to $\chi^2$): 100% DR, 18% FAR
- The paper highlights that the intensity of PD spikes increases with $\beta$, and PD with $\beta \ge 2.5$ shows a proportional increase with attack intensity.
Limitations and Future Work
Limitations: * The paper focuses primarily on SYN flooding attacks due to their prevalence, although the approach is stated to be applicable to other flooding types (TCP, UDP, ICMP). * The experimental analysis of the impact of Sketch parameters (width and depth) is referenced as being detailed in other work [6], but the specific optimal parameters used in this paper ($w=1024, d=5$) are presented as a trade-off between accuracy and memory. * The paper acknowledges that detecting low-intensity attacks is challenging, and while their method improves upon existing ones, the distinction between subtle attacks and normal traffic variations remains a core difficulty.
Future Work: * Developing defense mechanisms to trigger automatic reactions against detected ongoing attacks. * Investigating methods to reduce the amount of monitored data on high-speed networks. * Analyzing the impact of sampling on the precision of the Power Divergence measure.
Summary updated