Open RAN Federated Learning for Privacy-Safe Local Model Aggregation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing federated learning schemes in open radio access networks require excessive data transfer and resource consumption for training a global model, which can be costly and inefficient, especially when dealing with privacy-sensitive data.
Innovation Solution
Implement a federated learning method where local training is performed at Near-RT RICs or E2 nodes, with a service management and orchestration framework (SMO) configuring and aggregating trained local models to create a global model without transferring raw data, using O1 configuration management notifications.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Extent of automation
If federated learning is implemented with central server aggregation, then model training can be performed distributedly, but excessive data transfer and resource consumption occur
Solution Approach 1:
The patent extracts and processes only the essential model parameters and gradients rather than transferring complete training datasets. By extracting only the necessary information (model updates, not raw data), the system reduces data transfer volume and energy consumption while maintaining the federated learning benefits of distributed training.
Solution Approach 2:
The system creates and exchanges copies of model parameters and intermediate results between distributed entities. Instead of moving raw data, copies of trained model components are exchanged through the communication interface, reducing the amount of data transferred and the energy required for data transmission.
2Productivity
If raw data is transferred for centralized training, then comprehensive model training is achieved, but data privacy is compromised
Solution Approach 1:
The patent extracts only the necessary model information and training results from each distributed entity without transferring the raw data itself. By taking out only the essential parameters and gradients while leaving the original data local, the system maintains model training comprehensiveness while protecting data privacy.
Solution Approach 2:
The system introduces an intermediary communication interface that facilitates secure exchange of model parameters between distributed entities. This intermediary layer enables comprehensive model training through coordinated parameter updates without direct access to or transfer of raw data, thus protecting privacy while maintaining training effectiveness.
3Measurement precision
If extensive data exchange is performed for model aggregation, then accurate global model is obtained, but communication overhead increases
Solution Approach 1:
The patent extracts and transmits only the essential model parameters, gradients, and update information required for aggregation rather than exchanging complete datasets. This extraction approach maintains global model accuracy by preserving all necessary training information while significantly reducing communication overhead.
Solution Approach 2:
The system performs partial data exchange by transferring only the necessary portions of model information (parameters, weights, gradients) rather than complete datasets. This partial action approach achieves sufficient model accuracy through selective information exchange, reducing communication complexity and overhead.
Data Source
AI summary
A federated learning (FL) procedure is disclosed between a SMO/Non-RT RIC acting as FL aggregator and a Near-RT RIC/E2 node acting as a FL client. The SMO configures a plurality of Near-real time radio access network intelligence controllers (Near-RT RIC) or E2 nodes of the an open radio access network to perform local training of a machine learning local model available in each of the Near-real time radio access network intelligence controllers (Near-RT RIC) or E2 nodes and generate a trained machine learning local model. The trained machine learning local models are then aggregated by the SMO to generate a machine learning global model which is distributed to the Near-RT RIC/E2 nodes. Configuration of the local model in the Near-RT RIC/E2 nodes can be achieved via O1 configuration management notification.