A channel adaptive gradient compression method based on federated learning

Abstract

The invention discloses a channel adaptive gradient compression method under federated learning, which includes the following steps: S1, each terminal sends a known training sequence to the edge server through an uplink channel; S2, the edge server sends a known training sequence according to each terminal Estimate the uplink channel status of each terminal by the degree of degradation received after the training sequence passes through the channel, and store the channel gain parameters of each uplink channel; S3, the edge server uses the channel status of the terminal and the set total quantization bit number to solve the optimization problem, Allocation of the number of quantized bits; S4. The edge server broadcasts information on the number of quantized bits that can be used by each terminal participating in the federated learning; S5. The terminal finds the number of quantized bits that can be used by itself in the broadcast information, and quantifies the local update gradient. The invention has the advantages of effectively alleviating the influence of wireless communication channel degradation on federated learning training effect, alleviating communication bottleneck, improving system anti-noise performance and the like.

Description

technical field [0001] The invention belongs to the technical field of federated learning gradient compression, and in particular relates to a channel adaptive gradient compression method under federated learning with side-end cooperation. Background technique [0002] With the rise of mobile smart devices, a large amount of data is scattered at the edge, limited communication resources and user privacy protection make the distributed machine learning framework - federated learning develop rapidly. In the federated learning framework, the terminal uses local data for local model training, uploads the trained local model parameters to the edge server, and the edge server aggregates all local model parameters to update the global model parameters, and then broadcasts the global model parameters, and the entire federated learning is carried out. The above steps are iterated until the global model converges. However, due to the increasing complexity of user data, the parameter ...

AI Technical Summary

Problems solved by technology

[0004] 1. The terminal cluster is usually located within the service range of the same edge server, and its uplink channel can be regarded as multiple independent and orthogonal sub-channels divided by the total uplink channel, and the total channel capacity is limited

[0005] 2. The uplink channel of the terminal has channel fading and channel noise that are common in wireless communication channels. When the quantized stochastic gradient is transmitted through this type of fading channel, the data will be damaged to a certain extent.

When the edge server uses lossy stochastic gradients for aggregation and global iteration, the resulting global model will deviate from the ideal global model, making federated learning performance worse

Images