A Decentralized Federated Learning-Based Approach for Fault Detection in Optical Networks

Optical networks offer an ultra-high transmission capacity and serve various online applications (e.g., 5G, IoT, AR/VR, telemedicine). Preventing faults that cause packet losses or even link interruption becomes vital to ensure the reliability of these networks and, consequently, access to vital online services. Moreover, as the volume of telemetry data rapidly increases, data processing is often done in the cloud, which can open up breaches of unauthorized data access and raise concerns about scalability. Therefore, this work proposes a decentralized federated learning (FL)-based approach that exploits the principal component analysis (PCA) to perform confidentiality-preserving fault detection in optical networks. Unlike centralized FL-based approaches, the PCA is split into several local PCAs trained with subsets of the entire telemetry dataset. Thereafter, each local model exchanges its parameters in a peer-to-peer manner to learn the information extracted from their local data. As local PCAs are trained with only normal data (i.e., without faults), these models become sensitive to data that indicate anomalies, enabling the detection of faults. Moreover, a scrambling technique is applied to shuffle the order of the dataset, hiding the structural dependency among samples from malicious agents. Combining decentralized FL with the scrambling technique can enhance data confidentiality and cope with network scalability, as the processing of the dataset will be distributed over several nodes, hindering the access of malicious agents. Results on a testbed-derived dataset show no penalties for adopting the proposed disaggregated solution, i.e., the performance is the same as that of the centralized solutions.