Leveraging Binarised Neural Networks for SDC Control

Watch the video

Military / Coalition Issue

The success of many military missions heavily relies on the timely access and analysis of data that often come from different sources that can be widely distributed across the military network. On the one hand, the distributed nature of the data complicates their analysis which often forces network operators to adopt simple distributed mechanisms for network control that run based on local data. On the other hand, the analysis of the data is difficult by itself and when not possible or too time-consuming the network operators have nothing but to rely on simple heuristic policies or empirical rules to manage their networks and support their missions.

Core idea and key achievements

With the Software Defined Coalition (SDC) architecture proposed in the DAIS program, data like network traffic measurements, resource availability and mission states are gathered at one or multiple logically-centralized network entities, the SDC controllers. This centralization facilitates the analysis of data and the derivation of corresponding optimal network policies. In addition, the SDC controllers can use their available data and computation resources to train Machine Learning (ML) models to uncover hidden information in the data, predict otherwise unexpected events and improve their overall network operations. However, due to high network dynamics the controllers may be fragmented from the nodes they manage rendering impossible the access to the trained ML models to infer the network policies. An approach that is robust to network fragmentation events is therefore needed.

Key achievements include the development of:

  • The adoption of Binarised Neural Networks (BNN) to perform the inference of the ML model in a lightweight manner such that even resource-constrained mobile handheld devices can afford to run. Therefore, model inference is possible even when the controllers are fragmented from the rest of the network.
  • An extension of this ML architecture for collaborative training among multiple BNN models distributed in a network using the Federated Learning (FL) paradigm.
  • A proof-of-concept prototype implementation using the P4 Software Defined Network (SDN) programming language.

image info

Implications for Defence

The BNN and P4 SDN language together will allow the running of ML models everywhere in the military network even at lightweight handheld devices and this way make intelligent decisions for network control, resource allocation and dissemination of information according to the mission needs in a distributed, robust and flexible manner.

Readiness & alternative Defence uses

TRL 2/3. Software prototype based on the P4 language available.

Resources and references

Samples of related publications include:


Yale University, Imperial College, IBM US, ARL