An approach to predictively securing critical cloud infrastructures through probabilistic modeling

Satvik Jain, Arun Balaji Buduru, Anshuman ChhabraarXiv 2018 · arXiv

Introduction

As critical infrastructures—including banking, utilities, and military systems—increasingly migrate to cloud environments, they become prime targets for sophisticated, well-funded cyberattacks. Traditional security measures are largely reactive, responding to breaches after they occur. However, for critical systems, defenders require predictive capabilities to anticipate and mitigate threats before they materialize.

While existing research has explored predictive security using Game Theory, Machine Learning (ML), and Bayesian Networks, these approaches often possess significant limitations. Game theory frequently assumes perfectly rational attackers, standard ML models struggle with unforeseen patterns, and Markov Chains often fail to incorporate the probabilistic nature of human behavior. This paper addresses these gaps by proposing a framework based on Markov Decision Processes (MDPs) that captures both operational cloud behavior and the probabilistic actions of users to forecast vulnerable system states.

Methodology

The proposed framework models a cloud subsystem as an MDP to derive an optimal security policy. The process follows a structured pipeline designed to handle the complexities of real-time cloud data.

The predictive security framework pipeline, from data collection to policy deployment.

Feature Engineering and State Abstraction

The system state is defined by seven features across three primary categories: Normal User Behavior (e.g., HTTP request counts, unique users), Subsystem Performance (latency, response time), and Malicious Activity (presence of DoS attacks).

Initial discretization of these continuous features results in a massive state space of 51,200 states. To overcome the computational "state explosion" problem inherent in MDPs, the authors utilize unsupervised clustering (specifically K-Means) to abstract the environment into approximately 1,000 manageable "abstract states."

MDP Formulation and Policy Generation

The MDP is defined by actions (remaining in a state or transitioning) and a custom reward function tied to a "Risk Metric." This metric is a weighted sum of system features; if a state's risk exceeds a specific threshold, it is classified as "risky." The MDP rewards the policy for identifying these risky states, effectively generating a state transition tree where future vulnerabilities become identifiable leaf nodes.

Experimental Setup

To evaluate the framework, the authors simulated a critical cloud infrastructure by deploying a banking application on AWS Elastic Beanstalk.

  • Traffic Simulation: Normal traffic was generated using Apache JMeter (GET/POST requests), while malicious traffic was simulated using Hping3 to launch SYN, UDP, and ICMP flood attacks (Denial of Service).
  • Monitoring: Snort was employed to detect attacks and generate alerts. Data was collected over 300-second windows to construct the state features.

Key Results and Analysis

The framework demonstrated high efficacy in both computational efficiency and predictive accuracy. A critical finding was the impact of the chosen abstraction method on the model's ability to represent the original state space.

Comparison of clustering algorithms showing K-Means (KME) achieving peak accuracy at 1,000 clusters.

Performance Metrics

  • Predictive Accuracy: The framework achieved a peak accuracy of 96.51% in identifying risky states when mapped back to the original state space. As shown in the results, K-Means (KME) with a Euclidean distance metric significantly outperformed Gaussian Mixture Models (GMM), which only reached approximately 74% accuracy.
  • Computational Speed: Testing various MDP solvers revealed that Modified Policy Iteration was the most efficient, solving the MDP in just 0.033 seconds—orders of magnitude faster than standard Value Iteration.
  • Optimal Abstraction: The study found that 1,000 clusters provided the ideal balance between state reduction and accuracy. Furthermore, empirical testing indicated that a low discount factor ($\gamma = 0.1$) yielded the highest predictive accuracy for this specific environment.

Conclusion and Future Work

This research demonstrates that combining MDPs with unsupervised clustering can effectively predict risky states in cloud infrastructures without the computational bottlenecks typically associated with dynamic programming. By incorporating human behavior into the state features, the model provides a more holistic view of system vulnerability.

Future work aims to expand the threat model beyond DoS attacks to include stealthier exploits like SQL injection and Advanced Persistent Threats (APTs). Additionally, the authors propose a hybrid approach, combining these subsystem-level MDPs with system-wide Bayesian Networks to enable global breach prediction with minimal overhead.

Summary updated

BibTeX
@misc{jain2018approach,
  author = {Satvik Jain and Arun Balaji Buduru and Anshuman Chhabra},
  title = {{An approach to predictively securing critical cloud infrastructures through probabilistic modeling}},
  year = {2018},
  eprint = {1810.11937},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1810.11937},
}
Figure