Privacy-Preserving Federated Learning Framework for Open Radio Access Networks (ORAN)

Open Radio Access Network (ORAN) is considered the next-generation RAN, which enables several features such as network flexibility, interoperability, and cost efficiency. Leveraging Artificial Intelligence (AI) and Machine Learning (ML) techniques has become commonplace in ORAN applications. The modularized nature of the ORAN architecture and the limitations in traditional ML approaches intensify the requirement of Federated Learning (FL) for training ML models in ORAN environments. However, in multi-BS environments, the conventional plaintext model update sharing of FL is vulnerable to privacy breaches like inference and deep-leakage gradient attacks. Hence, our proposition introduces an innovative blockchain-based framework to conduct FL securely and protect privacy with the support of the Open Radio Access Network (ORAN). Our approach builds upon the traditional masking method for sharing model parameters and enhances it with novel features. These features include individual validation for BSs, the selection of distributed aggregators, and validation for final model aggregation. Our scheme facilitates the sharing of sensitive data among multiple BSs while bolstering privacy and adding security layers without compromising performance metrics. To assess the efficacy of our proposal, we implement the framework atop a Hyperledger Fabric blockchain. Furthermore, comprehensive formal and informal security analyses are conducted to demonstrate the robust and privacy-preserving nature.