Federated Learning as a Service: Enhancing Trustworthy AI in B5G/6G Networks

By Samudrapom DamReviewed by Susha Cheriyedath, M.Sc.Aug 2 2023

In an article in press with the journal Computer Communications, researchers proposed a novel federated learning (FL)-as-a-Service framework (FLaaS) that provides flexible mechanisms beyond fifth-generation wireless (B5G)/sixth-generation (6G) wireless network to allow the exploitation of FL services by end users.

Background

The next-generation mobile networks are expected to leverage advanced artificial intelligence (AI) methodologies to assist users in running novel services and applications and to achieve reliable and effective in-network functionalities. AI features will be integrated with mobile networks to improve several management tasks, such as virtual network function orchestration and user-plane operations, such as capability augmentation of mobile network user-run applications.

The automotive sector is expected to be substantially impacted by the integration of AI in mobile networks. For instance, vehicular applications depending on video streaming transmission over the mobile network can exploit AI to predict potential quality of experience (QoE) degradation based on quality of service (QoS) parameters to implement countermeasures to prevent such degradations, such as avoiding risky maneuvers.

Although several studies have investigated the feasibility of using machine learning (ML) techniques to optimize the accuracy of QoS-QoE prediction tasks, the current focus is primarily on increasing the trustworthiness of these models. B5G/6G networks will provide services to a substantial number of devices, each with advanced sensing capabilities, generating a very large amount of valuable user data. This user data can be used to develop AI models with higher accuracy through ML.

However, user data exploitation on the network side to feed training algorithms can lead to privacy concerns. Thus, decentralized approaches that address privacy concerns are necessary for B5G/6G networks.

FL to address the privacy concerns in B5G/6G networks

FL can become a suitable approach as it contains a learning paradigm that enables several parties, such as data owners, to learn an AI model collaboratively without disclosing private raw data. Specifically, the AI model training phase occurs locally, and models are shared with a central orchestrator instead of private data. The central aggregation entity develops a global model by aggregating the shared local models.

Thus, the FL paradigm can be crucial in cases where AI model training centralization is impossible due to computation overhead and intolerable communication. Although FL can address privacy issues to increase the trust of users in AI services, the explainability of AI systems is crucial to make AI trustworthy, which necessitates the development of FL approaches for eXplainable AI (XAI) models.

Enabling FL of XAI models in B5G/6G networks

In this paper, researchers proposed a comprehensive framework enabling FL services in future B5G/6G networks based on the as-a-service paradigm. The study aimed to exploit the synergic interface between next-generation wireless networks and trustworthy AI.

The proposed FLaaS framework can allow mobile users to explore the AI models that were made available by the network for different applications, such as for the prediction of QoE for video-streaming applications, to obtain such AI models from the B5G/6G network, and utilize them for inference tasks.

Additionally, the framework can provide protocols to mobile users to leave/join a federation and participate in the corresponding XAI model training by exchanging the global and local versions of the model with the centralized aggregation entity in the network.

The infrastructure for central coordination can be deployed at the network edge by exploiting European Telecommunications Standards Institute (ETSI) multi-access edge computing (MEC) architecture. Distributed functions, such as user training modules, can be offloaded as a dedicated application to the edge or placed at the user device to allow users with limited computational abilities to participate in FL operations.

Researchers exploited Takagi-Sugeno-Kang Fuzzy rule-based systems (TSK-FRBSs) to achieve explainability. The TSK-FRBSs are effective for dealing with regression tasks as they can achieve high modeling capability of complex systems and improve the model interpretability and performance in scenarios with some degree of noise due to the natural linguistic representation of numeric variables. The novel FLaaS framework supported the proposed federated-XAI (Fed-XAI).

Researchers validated their approach by performing an extensive experimental analysis. They evaluated the performance of the XAI model learned in a federated manner on a public QoS-QoE dataset. They also investigated the effect of system deployment and the underlying computation and communication infrastructure performance on the learning process through system-level simulations. Moreover, the FL scheme was also compared with two baselines, including local learning and centralized learning.

Significance of the study

The proposed FLaaS framework for FL of XAI models ensured the preservation of data privacy while enabling collaborative learning of AI models. Additionally, the AI models used in this study were explainable by design, allowing all entitled stakeholders to oversee and monitor a transparent decision-making process and addressing the trustworthiness issue.

Researchers successfully displayed the applicability of the FLaaS framework to a QoE forecasting service based on a vehicular networking use case. Specifically, they demonstrated the feasibility of leveraging FL of XAI models for the QoE forecasting task to increase the forecasting accuracy compared to local learning.

These results were confirmed by the experimental analysis performed on a simulated dataset. The deployment of the FLaaS components in edge-based environments significantly reduced the overall time required for federated model generation.