Preserve Privacy‐HD
Aswathy Ravikumar, Harini Sriraman · 2023
A distributed software framework called Preserve Privacy-HD trains deep neural network models in numerous decentralized edge systems. Every edge system will have a data-parallel model and provide the model with incoming input. In each edge device, the model is run concurrently. They will send an asynchronous message to the multi-parameter server including the local weights computed by each edge device. The model is updated by the multi-parameter server, which determines the global weights based on the local weights. The edge device then downloads the model for additional processing. Healthcare data are not shared with other devices, thanks to Preserve Privacy-HD. Instead, only the weights are shared to adjust the model. The suggested framework's metrics for accuracy, training duration, and privacy have all been confirmed. Maintaining the original client data on local systems protects their privacy while leveraging the server in subsequent layers and using all information from each system during training increases learning performance.