Detecting Behavioral and Structural Anomalies in MediaCloud Applications

Guido Schwenk, Sebastian BacharXiv 2014 · arXiv

Introduction

In the evolving landscape of cloud computing, platforms like the Alcatel-Lucent Media Cloud represent a significant shift toward high-performance, real-time interactive media services. These environments are characterized by extreme dynamism; "atomic components"—minimized virtual machines—are frequently re-mapped across physical servers to optimize latency. Such fluidity renders traditional, signature-based intrusion detection systems (IDS) ineffective. This paper presents a dual-layered security framework designed to detect anomalies by monitoring both the behavioral metrics of individual components and the structural integrity of the application’s communication graph.

Methodology

The proposed system employs an unsupervised learning approach, utilizing One-Class Support Vector Machines (OCSVM) to define "normality" without requiring labeled attack data. The detection is split into two distinct modules: behavioral and structural.

Behavioral Anomaly Detection

Behavioral detection focuses on the internal "vitals" of individual software components. The system tracks five specific metrics that characterize the resource consumption and communication patterns of a component.

Table 1: Description of the five behavioral metrics ($V_t$) monitored for each software component.

To account for the temporal nature of these metrics, a sliding window approach is used. This transforms the raw data into a higher-dimensional trajectory, which is then processed by an OCSVM with a Radial Basis Function (RBF) kernel. This creates a hypersphere boundary around normal behavior; any data point falling outside this boundary is flagged as a potential threat.

Structural Anomaly Detection

While behavioral detection monitors individual nodes, structural detection analyzes the application's communication topology. The researchers utilize a "Bag of Substructures" (BoS) approach, decomposing the application graph into 1st-degree neighborhoods.

A key innovation in this module is the use of sum-decomposable kernels (such as Linear or Histogram Intersection kernels). Unlike standard graph kernels that provide a binary anomaly score, these kernels allow the final score to be mathematically decomposed. This enables the localization of anomalies, providing administrators with a visual heatmap to identify the specific nodes or edges responsible for the alert.

Experimental Evaluation

The framework was evaluated using logs from a real-world video conferencing application (up to 160 components) and synthetic datasets designed to simulate sophisticated attacks. Since real-world logs lacked malicious activity, the authors crafted anomalies using parametric sigmoid functions for behavioral shifts and manual graph edits for structural attacks, such as traffic hijacking.

Key Results

The system demonstrated high efficacy across both detection layers:

  • Behavioral Accuracy: The system achieved a 100% True Positive Rate (TPR). On static data, this was reached with a window size of $s=60$, while more dynamic data required a larger window of $s=160$ to maintain accuracy.
  • Structural Accuracy: The structural module achieved perfect graph-wise prediction (1.0 AUC) on recorded data. Even in complex synthetic scenarios with mixed anomalies, the AUC remained high at approximately 0.98.
  • Operational Efficiency: In optimal configurations, the False Positive Rate (FPR) remained below 1%, a critical threshold for minimizing alert fatigue in production environments.

Conclusion

By combining component-centric behavioral monitoring with graph-based structural analysis, this framework provides a robust defense mechanism for dynamic media clouds. The ability to localize threats within complex communication graphs represents a significant advancement over "black box" detection methods. Future work will explore deeper neighborhood analysis and strategies to mitigate model poisoning in online learning scenarios.

Summary updated

BibTeX
@misc{schwenk2014detecting,
  author = {Guido Schwenk and Sebastian Bach},
  title = {{Detecting Behavioral and Structural Anomalies in MediaCloud Applications}},
  year = {2014},
  eprint = {1409.8035},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1409.8035},
}
Figure