Confidentiality-preserving real-time localization of soft failures in optical networks based on PCA and MLaaS

Abstract

Proactive management of soft failures is crucial for enhancing the reliability of optical networks. However, developing solutions that are simultaneously accurate, operate in real time, ensure data confidentiality, and scale effectively represents a significant challenge. This paper proposes a method for soft failure localization that ensures data confidentiality. The approach is devised for a scenario where the data owner (e.g., the network provider) elaborates its confidential data (e.g., telemetry data) through machine learning services provided by a third party (i.e., machine learning as a service). Data confidentiality and, as an important by-product, reduced data exchange are achieved by using principal component analysis-based data dimension reduction before transmission. The data are then sent to a third party, where they are processed using a semi-supervised K-means clustering algorithm. The resulting cluster labels are returned to the data owner, who performs label matching to localize potential failures. The method’s effectiveness is validated in terms of failure localization accuracy, achieving up to 98.5% on large-scale simulated datasets and 98% on small-scale experimental data.