Efficient Forwarding Anomaly Detection in Software-Defined Networks

Qi Li, Yunpeng Liu, Zhuotao Liu, Peng Zhang, Chunhui PangIEEE TPDS 2021 · DOI

Abstract

Software-Defined Networking (SDN) promises centralized control, but discrepancies between the control plane's intent and the data plane's actual behavior—known as forwarding anomalies—can lead to severe security and performance issues. We present FADE and its scalable extension, iFADE, efficient systems designed to detect and localize traffic hijacking and interception with high accuracy and minimal overhead. By optimizing flow selection and probe placement, these systems achieve over 95% detection rates while reducing control message overhead by up to 90% compared to existing solutions.

Introduction

In Software-Defined Networking, the control plane (the "brain") dictates traffic handling to the data plane switches (the "brawn"). When the data plane fails to execute these instructions correctly due to hardware glitches, software bugs, or malicious attacks, the network suffers from forwarding anomalies. These issues, ranging from simple packet drops to sophisticated traffic interception, can compromise network integrity. Detecting these anomalies efficiently, without introducing significant latency or control plane congestion, remains a critical challenge in SDN management.

The Forwarding Anomaly Problem

Forwarding anomalies typically manifest in two forms:

  1. Traffic Hijacking: Packets are dropped or diverted to incorrect destinations, failing to reach their intended target.
  2. Traffic Interception: Packets are detoured through unauthorized middleboxes but eventually return to their correct path. This "looping" behavior makes interception significantly harder to detect than simple hijacking.

Existing detection methods often struggle to balance accuracy, overhead, and coverage. Probing-based approaches can saturate the controller, while statistical analysis often requires monitoring every flow, which is unscalable in large-scale environments.

The FADE and iFADE Architecture

The researchers proposed a two-tier architecture to address these challenges by shifting from exhaustive monitoring to strategic, minimalist measurement.

The FADE system architecture and workflow, from flow selection to anomaly localization.

FADE: Minimalist Detection

FADE (Forwarding Anomaly Detection architEcture) focuses on minimizing the monitoring footprint through three core strategies:

  • Minimal Flow Selection: Instead of monitoring every rule, FADE uses a reversed Depth-First Traversal (DFT) on the network's forwarding graph to identify the smallest set of flows that covers all active forwarding rules.
  • Optimized Probe Placement: Using probabilistic modeling, FADE determines the minimum number of monitoring points required along a path to detect anomalies effectively, avoiding the need for hop-by-hop probing.
  • Synchronized Passive Measurement: FADE installs "passive" measurement rules. Ingress rules tag existing traffic with unique labels (e.g., in ToS or VLAN fields), and downstream rules count these labeled packets. Discrepancies in counts between monitoring points signal an anomaly.
Network topology (b) and its corresponding forwarding graph (c) used for minimal flow selection.

iFADE: Scalable Optimization

To support large-scale data centers, iFADE introduces two key optimizations:

  • Flow Aggregation: It groups flows with identical paths and actions, allowing a single measurement rule to monitor multiple flows simultaneously.
  • Round-based Scheduling: To accommodate switches with limited memory (TCAM), iFADE uses an Integer Linear Programming (ILP) heuristic to schedule detection tasks across multiple rounds, ensuring full coverage over time without exceeding hardware limits.

Evaluation and Results

The systems were evaluated using a prototype implemented in 12,000 lines of Java code, tested against realistic topologies from the Topology Zoo and large-scale Forwarding Information Base (FIB) data from Internet2. The physical testbed utilized the Floodlight controller and Open vSwitch.

Key findings include:

  • High Detection Accuracy: Both FADE and iFADE achieved a True Positive Rate (TPR) over 95% and a True Negative Rate (TNR) near 100%.
  • Minimal Performance Impact: Data plane throughput degradation was kept below 3%, ensuring that monitoring does not hinder production traffic.
  • Superior Scalability: iFADE reduced the required measurement rules by over 40% compared to FADE. Furthermore, iFADE generated 90% fewer control messages than SPHINX, a leading statistics-based detection system.

Conclusion and Future Work

FADE and iFADE demonstrate that comprehensive rule coverage can be achieved with a fraction of the overhead typically associated with SDN monitoring. Future research will address compound anomalies where multiple types of attacks overlap, explore label space optimizations for networks with high Quality of Service (QoS) demands, and refine scheduling heuristics for extremely high-churn environments.

Summary updated

BibTeX
@article{li2021efficient,
  author = {Qi Li and Yunpeng Liu and Zhuotao Liu and Peng Zhang and Chunhui Pang},
  title = {{Efficient Forwarding Anomaly Detection in Software-Defined Networks}},
  journal = {IEEE Trans. Parallel Distributed Syst.},
  year = {2021},
  doi = {10.1109/tpds.2021.3068135},
}
Figure