Design Considerations and Guidelines Product Optimization by Original Equipment Manufacturers ( OEM )
This is a typical cross-silo FedL setting where multiple organizations in possession of equipment provided by an OEM , are willing to participate in a collaborative ecosystem to improve the performance of the equipment but stipulate that they cannot sharing raw equipment data with each other or with the OEM . As the variety and veracity of the equipment data gathered from one manufacturing site is not enough to mine complex data patterns , the objective is to gather the data around the equipment operating in different working conditions at their corresponding organizational sites and perform a collaborative model training using FedL .
• Type of clients and mode of operation : Data gathering associated with operational efficiency of equipment is typically collected in parallel to equipment ’ s normal operational tasks and thus the product ( equipment ) fine-tuning / optimization does not demand real time data analytics . As per the decision model , the type of clients in this use case can be deployed either as on-premises edge servers if the global ML model is not computationally demanding or on cloud servers to support a computationally demanding model training task . This decision is driven by the amount of training data and frequency of model exchanges between participatory organizations .
• Global model training orchestrator : In such cross-organizational settings , participatory organizations may be hesitant in letting a single organisation orchestrate the global model training . As per the decision model , a decentralized orchestration is suited for this use case where a third party ( chosen by participatory organizations ) or the OEM itself can become the orchestrator . The OEM can also employ an interesting business model where the cost of model training orchestration can be accommodated by receiving a copy of the trained global model . It may benefit OEM to produce / supply more optimized products in the future . Current third-party ( apart from OEM ) solutions in this direction provide a secure model / data sharing platform to be utilized by participatory clients during model exchanges .
• Security provisions : The use case is a typical cross-silo FedL setting in which participatory organizations cannot afford to expose even their trained model parameters to each other or to the orchestrator . Thus , additional security provisions are needed in this use case to prevent any sensitive data / model leakage . In addition to that , a more secure data-sharing platform must be utilized to keep transactional ( model exchanges ) records intact during and post model training .
• Technological enablers : Sophisticated deep learning models and computationally rich model training platforms are key enablers for this use case . In addition to that , DLT such as Blockchain and model encryption techniques such as differential privacy and homomorphic encryption should be utilized for secure and auditable model exchanges .
IIC Journal of Innovation 31