Federated Model Update Serialization in Communication Networks
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Collating training data for machine-learning models in communication networks at a single network entity is slow, resource-intensive, and poses security risks, especially when dealing with sensitive data.
Innovation Solution
Implementing a collaborative (federated) learning process where local model updates are trained at multiple network entities and combined at an aggregator entity, using serialized representations that omit model structure information to ensure data privacy and reduce resource overhead.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If training data is collated at a single network entity, then machine-learning models can be trained centrally, but the process becomes slow and resource-intensive
Solution Approach 1:
The patent segments the centralized data collation process into distributed local training operations at multiple network entities. Each entity trains models locally using its own data, eliminating the time-consuming process of gathering all data at one location while maintaining model training capability through distributed computation.
Solution Approach 2:
The patent transitions from a single-dimension centralized training approach to a multi-dimensional distributed training architecture. By adding the spatial dimension of multiple network entities performing training simultaneously, the system reduces overall training time while maintaining model quality through aggregation of local model updates.
2Measurement precision
If training data is collated at a single network entity, then centralized model training can be performed, but resource consumption increases
Solution Approach 1:
The patent divides the resource-intensive centralized training task into smaller local training operations distributed across multiple network entities. Each entity consumes resources locally for training, avoiding the集中 resource consumption at a single entity, thereby reducing overall network resource burden while achieving the same model training objective.
Solution Approach 2:
The patent creates copies of the model at multiple network entities, allowing each to perform local training independently. This eliminates the need for one entity to handle all training resources, distributing the computational burden across multiple copies of the model instance rather than concentrating resources at a single location.
3Measurement precision
If training data is collated at a single network entity, then centralized model development can occur, but security risks increase for sensitive data
Solution Approach 1:
The patent segments the sensitive training data across multiple network entities, preventing concentration of sensitive information at a single location. Each entity retains its data locally, reducing security risks associated with data aggregation while still enabling collaborative model training through sharing of model updates rather than raw data.
Solution Approach 2:
The patent extracts only the necessary model update information from local training operations and transmits this to the aggregating entity, leaving sensitive raw training data at local entities. This extraction approach maintains model training capability while removing the security vulnerability of transmitting or storing sensitive raw data at centralized locations.
Data Source
AI summary
A method performed by a first network entity in a communications network includes training a model to obtain a local model update including an update to values of one or more parameters of the model, in which training the model includes inputting training data into a machine learning algorithm. The method further includes applying a serialisation function to the local model update to construct a serial representation of the local model update, thereby removing information indicative of a structure of the model, and transmitting the serial representation of the local model update to an aggregator entity in the communications network.