Design Considerations and Guidelines
On the other hand , in vertical FedL , data across multiple clients overlap on sample IDs but not on features . This is often used when clients participate to share the missing / intermediate features of the dataset . An example of cross-silo-vertical FedL in smart-manufacturing is monitoring the health degradation of an industrial bearing . This can be done by capturing images in a time-series and / or by recording parameters such as vibration , temperature etc .
Cross-device FedL is the most used paradigm of FedL where clients (> 10,000 ) are un-reliable , state-less , and usually participate in a horizontal FedL of a light-weight predictive model . An example of cross-device FedL for smart-manufacturing is the collaboration of similar assets across multiple sites of organizations where agile edge analytics is important . The connected edge devices form network clusters i . e ., all edge devices in a single cluster possess IID data about the same type of asset .
Digital twins can be the potential technology drivers for cross-device FedL as the concept provides a basis to make the transition from standalone , relatively unintelligent systems to a network of “ intelligent ” objects on the internet , facilitating and fueling the development of new value-added services enabled by access to data / model extracted from distributed physical assets .
Resource-constrained IoT edge devices are used as FedL clients if the global model is lightweight in terms of the number of model parameters and training data size . To validate this , the authors conducted experiments on single board computers ( Raspberry pi ) with 2GB RAM . A lightweight artificial neural network ( ANN ) with 2-hidden layers was used as a global model to be trained among three Raspberry pi devices and a centralized orchestrator ( Laptop machine ). The experiment results showed that almost 60 % of the memory was occupied during the model training process with > 85 % CPU utilization throughout the experiment . This shows that such resource constrained IoT edge devices will not be able to train a more complex ML model such as deep learning , for which the edge / cloud servers are the appropriate choices .
On the other hand , the clients involved in cross-silo FedL are limited ( 1-100 ), reliable , state-full and equipped with enough computational resources . Interoperability and complex data ownership structures remain a challenge in the implementation of both cross-device and crosssilo FedL , as such the use of standards is encouraged . In the case of manufacturing standards such as OPC UA and Asset Administration Shell ( AAS ) are emerging as key approaches to support information modelling that can be common across all participants in the FedL network . IEEE standard 3652.1™-2020 defines an architectural framework and requirements for the application of FedL . It sets out to act as a guide to promote the use of distributed data sources and FedL without violating regulations or ethical considerations .
In cross-silo FedL , any of the organisations can be elected as a trusted orchestrator and can manage the global model training task . However , when competing organisations are involved , electing a single orchestrator is often un-desirable considering the cost of manufacturing data
IIC Journal of Innovation 27