Using Online Planning and Acting to Recover from Cyberattacks on Software-defined Networks

Sunandita Patra, Alexander Velazquez, Myong H. Kang, Dana S. NauAAAI 2021 · DOI

TL;DR

Researchers have developed ACR-SDN, an automated system that uses AI-driven planning and acting to recover Software-Defined Networks (SDNs) from cyberattacks. By combining a refinement-based acting engine (RAE) with a Monte Carlo-based planner (UPOM), the system can autonomously diagnose issues and execute complex recovery procedures, reducing recovery time by approximately 32% compared to non-planning approaches.

The Challenge of SDN Security

Software-Defined Networking (SDN) is a modern architecture that centralizes network control, making networks more flexible but also creating a high-value target for attackers. When an SDN is hit by a "PACKET_IN" flooding attack or a controller failure, the speed and volume of the onslaught often outpace human administrators.

Existing defense systems are frequently manual or ad-hoc. While Machine Learning (ML) and Reinforcement Learning (RL) have been proposed for defense, they often require massive amounts of training data that are hard to obtain for niche cyberattacks. The challenge is creating a system that can leverage expert cybersecurity knowledge to act autonomously and choose the most effective recovery strategy in real-time.

The ACR-SDN Framework

The authors introduce ACR-SDN, which integrates two core AI components into the SDN management plane to automate recovery processes.

The multi-layered architecture of ACR-SDN, illustrating the integration of the RAE+UPOM planning engine within the management layer.

As shown in the architecture above, the system operates across three layers:

  1. RAE (Refinement Acting Engine): An "actor" that executes hierarchical refinement methods. These methods are sophisticated procedures written by experts that include programming constructs like loops and if-else statements. For example, a method might specify: "If the controller is overwhelmed, first migrate critical hosts to a new switch, then reboot the primary controller."
  2. UPOM (UCT-like Procedure for Operational Models): A "planner" that helps RAE decide which method to use. Since there might be multiple ways to recover, UPOM performs Monte Carlo rollouts—simulated "what-if" scenarios—to predict which recovery path has the highest probability of success and the lowest cost in the current environment.

The system uses the Web Application Messaging Protocol (WAMP) for asynchronous communication between the security manager, the AI planner, and the network components.

Experimental Methodology

The researchers developed a simulated SDN environment (using LibVirt, SDNCTL, and SysMon) to generate a test suite of 300 recovery tasks. These tasks covered three major attack classes: controller memory exhaustion, switch memory exhaustion, and controller-switch disconnection. Simulations scaled from 1 to 4 controllers and 16 to 64 switches to test the system's robustness.

Key Results and Performance

The authors evaluated ACR-SDN using five performance metrics, comparing the RAE+UPOM system against a purely reactive baseline. The results demonstrated that even a small amount of look-ahead planning significantly improves the efficiency and success rate of expert-defined procedures.

Impact of UPOM planning time on the success ratio of network recovery tasks.
  • Recovery Speed: Using the UPOM planner led to a ~32% decrease in the estimated time for recovery compared to reactive acting.
  • Success Ratio: As seen in the figure above, planning increased the overall success rate of recovery by 7%. The success ratio stabilizes quickly, showing that even 200ms of planning time provides a substantial boost over the reactive (0ms) baseline.
  • Cost-Effectiveness: The system achieved a 31% improvement in cost-effectiveness, balancing time-efficiency and success probability.
  • Retry Ratio: The number of failed attempts (retries) dropped significantly as the planner was given more time to simulate outcomes.

Conclusion and Future Directions

ACR-SDN demonstrates that combining expert knowledge (via RAE) with automated planning (via UPOM) creates a powerful defense mechanism for SDNs. By encoding human expertise into executable methods, the system remains predictable and auditable while achieving superhuman response speeds. Future work will focus on addressing the "knowledge engineering bottleneck" by simplifying how experts write refinement methods and integrating operational constraints to prevent security actions from accidentally disrupting mission-critical services.

Summary updated

BibTeX
@inproceedings{patra2021using,
  author = {Sunandita Patra and Alexander Velazquez and Myong H. Kang and Dana S. Nau},
  title = {{Using Online Planning and Acting to Recover from Cyberattacks on Software-defined Networks}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2021},
  doi = {10.1609/aaai.v35i17.17806},
}
Figure