Secure Collaborative Learning for Predictive Maintenance in Optical Networks.

Building a reliable and accurate machine learning (ML) model is challenging in optical networks when training datasets are business-sensitive. We propose a framework of secure collaborative ML learning for predictive maintenance on cross-vendor datasets. Our framework is based on federated learning and multi-party computation technologies. Each vendor builds a local ML model based on its own private data. A server builds a global ML model by aggregating multiple local ML models in a private-preserving way. The server computes only the sum of the local models but cannot see any local model individually by the multi-party computation technique. The vendor-confidential dataset is never exposed to the server or other vendors. Moreover, after the global ML model is deployed in optical networks, the measured data compared to the prediction are privately distributed to the local model owners, which is beneficial to vendors. We applied our framework to the remaining useful life (RUL) prediction of laser device. Our experiments show that an accurate ML model can be built using sensitive datasets in a federated learning setting.