Online Federated Learning for Real-Time Network Analysis

Communication networks are heterogeneous, distributed, and often lack centralized raw training data, making traditional Machine Learning (ML) difficult to adopt. Although Federated Learning (FL) addressed some of these drawbacks, FL, just like traditional (offline) ML, assumes that local data arrives in a statistically known manner, which makes them unable to adapt to the evolving statistics in communication patterns. To address this gap, we present a novel paradigm called Runtime Online Federated Learning (ROFL) for communication networks by modifying the FL paradigm into an online variant using continual/incremental learning principles of Online Learning (OL) to achieve just-in-time results i.e. in “runtime”. Specifically, we make the following contributions. (1) Our ROFL adopts an asynchronous aggregation for Global Model updates and adaptive activation rates and update periods for each node. (2) Unlike existing federated online approaches, ROFL is designed for previously unknown data to be available at runtime. We implement functionalities to listen for incoming packets at runtime, preprocess them and feed into the learning models. (3) As part of ROFL, we present a lighweight traffic classification model based on Natural Language Principles (NLP). When designed as a local model for each local node in ROFL paradigm, our classifier was able to seamlessly classify packets at runtime and continued to adapt to newer traffic in online fashion. (3) We evaluate our framework by emulating a real production-level network communication scenario. We show that our proposed framework is not only able to classify packets at runtime, but can also learn/train continuously once deployed, induce minimal communication overhead, and adapt to varying communication patterns. With modifications to the Local Model, our framework can be theoretically extended to any runtime federated task(s) which need to learn from continuous distributed data to make runtime inferences.