Abstract
Distributed Denial of Service (DDoS) attacks represent a significant threat to Internet security, particularly in cloud and big data environments where early-stage detection is notoriously difficult. Traditional methods often rely on single features or static parameters that fail to capture the nuanced, non-linear nature of modern network traffic. This paper proposes an Adaptive DDoS Attack Detection Method (ADADM) based on Multiple Kernel Learning (MKL). By defining five distinct traffic features and employing an ensemble learning framework, ADADM adaptively adjusts feature weights to maximize inter-class separation and minimize intra-class variance. Utilizing a sliding window mechanism to coordinate two specialized Simple Multiple Kernel Learning (SMKL) models, the proposed method achieves high detection accuracy and an exceptionally low false alarm rate in early-stage attack scenarios.
Introduction
Detecting a DDoS attack at its peak is straightforward due to the sheer volume of traffic; however, identifying an attack in its "early stage"—when zombie machines are just beginning to ramp up—is a complex challenge. In modern cloud environments, early-stage attack traffic often mimics legitimate user behavior, characterized by bidirectional communication and bursty patterns. Static detection models struggle with this stochastic behavior, often leading to high false alarm rates or missed detections. To address this, the ADADM framework introduces a multi-dimensional, adaptive approach designed to distinguish subtle anomalies from background noise.
The ADADM Framework
1. Multi-Dimensional Feature Engineering
To capture the diverse characteristics of network flows, ADADM utilizes five specialized features:
- Address Correlation Degree (ACD): Analyzes the distribution of source IP addresses and traffic burstiness.
- Flow Feature Value (FFV): Exploits the asymmetry in packet flows, which is particularly effective at identifying early-stage attacks where bidirectional flows are still present.
- Interaction Behavior Feature (IBF): Measures the level of interactivity between hosts to identify automated, non-human patterns.
- Multi-Feature Fusion (MFF): Integrates various behavioral characteristics to ensure model stability across different network conditions.
- Half Interaction Anomaly Degree (HIAD): Focuses on aggregated flows that are heavily mixed with background traffic noise.
2. Multiple Kernel Learning (MKL) Ensemble
Rather than relying on a single Support Vector Machine (SVM) kernel, which may not suit all data distributions, ADADM uses an ensemble of two specialized SMKL models:
- M-SMKL (Inter-class Mean Squared Difference Growth): This model uses gradient ascent to maximize the distance between "normal" and "attack" data points, forcing a clear separation between classes.
- S-SMKL (Intra-class Variance Descent): This model uses gradient descent to minimize the variance within each class, tightening the clusters to reduce overlap.
3. Adaptive Weighting and Sliding Window
During the training phase, the weights for each of the five features are adjusted dynamically. This allows the system to prioritize the most relevant indicators based on the current traffic profile. A sliding window mechanism then acts as a coordinator, synthesizing the outputs of both M-SMKL and S-SMKL to produce a final, robust prediction.
Experimental Evaluation
Dataset and Simulation
The researchers evaluated ADADM using the CAIDA "DDoS Attack 2007" dataset, comprising 21 GB of anonymous TCP traffic. To test the model's resilience, the team simulated three distinct traffic scenarios by applying random scaling factors to the features:
- Scaling: General adjustment of traffic volume.
- Narrowing: Reducing attack intensity to simulate stealthier incursions.
- Amplifying: Increasing background noise to test the model's precision.
Results and Discussion
The experimental results demonstrate that ADADM outperforms standard SVMs and other baseline methods (such as SimpleMKL and Nezhad et al.) across several key metrics:
- Detection Rate (DR): ADADM maintained a consistent detection rate of 78-79% across all simulated scenarios.
- False Alarm Rate (FR): Most notably, the model achieved an exceptionally low FR of 0.01%. In comparison, baseline methods saw false alarm rates spike as high as 70% when faced with fluctuating reference points.
- Total Error Rate (ER): ADADM consistently yielded the lowest total error rate (approximately 12%), proving its reliability in distinguishing early-stage attacks from legitimate traffic bursts.
Conclusion and Future Work
ADADM provides a sophisticated solution for early-stage DDoS detection by combining multi-feature fusion with adaptive ensemble learning. By balancing class separation with cluster concentration, the method achieves high precision with minimal false positives.
While effective, the authors note that the current weight adaptation process is not yet a convex optimization problem, which can impact convergence efficiency. Future research will focus on transforming the weight adjustment into a convex problem and testing the framework's efficacy in emerging domains such as the Internet of Things (IoT) and Cloud Robotics, where traffic patterns are increasingly unpredictable.
Summary updated