TL;DR
This paper introduces a novel, zero-training framework for detecting and mitigating DDoS attacks in decentralized Software-Defined Networking (dSDN). By leveraging lightweight port-level traffic statistics and a pre-trained Large Language Model (LLM) called DeepSeek-v3, the system achieves near-perfect detection accuracy with minimal computational overhead, offering a proactive and scalable defense against evolving threats.
What Problem Does This Paper Address?
Traditional centralized Software-Defined Networking (cSDN) architectures, while offering programmability, are vulnerable to single points of failure and scalability issues, especially under attack. Decentralized SDN (dSDN) aims to solve these by distributing control, but this distribution introduces new complexities and attack surfaces. DDoS attacks can overwhelm local controllers, disrupt synchronization, and cascade failures across domains. Existing DDoS detection methods often rely on complex flow-level features, require extensive training, and are not well-suited for the dynamic and distributed nature of dSDN. This paper addresses the critical need for a lightweight, adaptive, and training-free DDoS detection and mitigation solution specifically designed for dSDN environments.
Methods
The core of the proposed solution lies in a novel integration of several techniques:
- Decentralized SDN (dSDN) Testbed: The authors implemented a dSDN testbed using Mininet for network emulation and the Ryu controller for programmability. This setup features multiple local controllers, each managing its domain, and utilizes Multiprotocol Label Switching (MPLS) for strict source routing to ensure deterministic traffic forwarding.
- Port-Level Feature Extraction: Instead of computationally expensive flow-level statistics, the system relies on lightweight port-level counters: received packets, received bytes, transmitted packets, and transmitted bytes. These are aggregated every 10 seconds, providing a balance between responsiveness and stability.
- Zero-Training LLM Inference with DeepSeek-v3: The key innovation is using the pre-trained DeepSeek-v3 LLM without any fine-tuning or retraining. The aggregated port-level features are converted into natural language prompts. These prompts are augmented with examples of benign traffic to enable "in-context learning." The LLM then dynamically infers whether the current traffic pattern is normal or indicative of a DDoS attack.
- Integrated Detection and Mitigation: Upon detecting a potential DDoS attack, the local controller immediately installs "drop rules" on the attacker's port, suppressing malicious traffic at its source before it can propagate. An automatic recovery mechanism re-enables the port after 30 seconds of inactivity.
Datasets
The experiments utilized a custom dataset generated from the dSDN testbed. This dataset consists of 25,933 prompts, each representing the traffic state of a specific port over a 10-second window. The dataset includes:
- Normal traffic (19,504 prompts)
- TCP Flood (2,151 prompts)
- UDP Flood (2,132 prompts)
- ICMP Flood (2,146 prompts)
Ground-truth labels were assigned based on the known timing and origin of the simulated attacks.
Key Insights and Ideas
- LLMs as Zero-Training Detectors: The paper demonstrates that pre-trained LLMs, when guided by well-crafted prompts and in-context learning, can achieve high-performance DDoS detection without the need for task-specific training. This significantly reduces computational costs and deployment time.
- Port-Level Statistics Suffice: The research highlights that simple, lightweight port-level statistics are sufficient for effective DDoS detection, challenging the notion that complex flow-level features are always necessary. This is crucial for the resource-constrained environments of local controllers in dSDN.
- Proactive, Source-Based Mitigation is Key: By integrating detection and mitigation directly at the local controller and attacker's port, the framework enables immediate, source-based defense, preventing attacks from spreading and minimizing their impact.
- dSDN's Unique Challenges Require Tailored Solutions: The study underscores that dSDN's distributed nature, while offering benefits, necessitates security solutions that are also distributed, adaptive, and can operate with low latency, a gap that this framework aims to fill.
Key Results
The proposed DeepSeek-v3 based framework achieved outstanding results:
- Near-Perfect Detection:
- Accuracy: 99.99%
- Precision: 99.97%
- Recall: 100.00%
- F1-Score: 99.98%
- AUC: 1.00
- Low False Alarm Rate: Only two benign instances were misclassified as attacks out of 25,933 prompts.
- Real-time Performance: In real-time experiments, mitigation was enforced within seconds of attack detection, rapidly reducing packet-in rates and stabilizing CPU usage of local controllers.
- Robustness Across Attack Types: The framework effectively handled TCP, UDP, and ICMP flood attacks, demonstrating its versatility.
- Superiority to Other LLMs: Comparative analysis showed DeepSeek-v3 outperforming other state-of-the-art LLMs (Gemma3, Qwen2, Mistral, LLaMA, Phi4) in zero-training DDoS detection for dSDN.
Limitations and Future Work
- Limitations:
- The study focused on flood-based DDoS attacks. The effectiveness against more sophisticated or stealthy attack types remains to be explored.
- The experiments were conducted in an emulated dSDN environment. Real-world deployment might introduce additional complexities.
- Future Work:
- Evaluate the framework on larger and more complex network topologies to assess scalability.
- Investigate the detection and mitigation of a wider range of DDoS attack types beyond floods.
- Test the framework in real-world environments with heterogeneous devices, dynamic traffic, and multi-tenant services for more realistic validation.
Summary updated