| Resumo: | The fifth-generation (5G) of broadband cellular networks is giving rise to new paradigms of distributed computing, such as Edge Computing and Multi-access Edge Computing (MEC). The possibility of hosting Machine Learning (ML) applications close to the end-users presents advantages, such as better privacy (e.g., sensitive data is not shared to other systems), the reduction of communication latency, improvement of application performance, and more efficient energy consumption. However, the Edge Computing and MEC paradigms also pose challenges to ML. For instance, the data can be distributed among distinct edges and might not be shared (e.g., due to privacy issues). Also, the ML models might be trained on edge devices with limited computational resources. In this paper, we propose a Federated ML architecture to train ML models on the 5G Edge, using decentralized data and light ML training algorithms. Our architecture includes edge nodes to train models with local data and a centralized node to aggregate the resulting models. As a case study, we address an International Revenue Share Fraud (IRSF) task, assuming a real-world dataset collected from a leading provider of analytics solutions for the Telecom industry. We evaluate our architecture during two iterations of a Federated ML procedure and then we compare it with a centralized baseline ML model that is currently adopted by the software company. Overall, the experimental results show that the proposed Federated ML approach outperforms the baseline ML model, thus supporting its potential usage to detect IRSF on the 5G mobile network edge.
|
|---|