Toward Securing Federated Learning Against Poisoning Attacks in Zero Touch B5G Networks

The zero Touch Management (ZSM) concept in 5G and Beyond networks (B5G) aims to automate the management and orchestration of running network slices. This requires heavy usage of advanced deep learning techniques in a closed-loop way to auto-build the suitable decisions, enabling to meet network slices' requirements. In this context, Federated Learning (FL) is playing a vital role in training deep learning models in a collaborative way among thousands of network slice participants while ensuring their privacy and hence network slice isolation. Specifically, running network slices may share only their model parameters with a central entity, e.g., Inter Domain Slice Manager, to aggregate them and build a global model. Thus, the central entity does not directly access the training data. However, FL is vulnerable to poisoning attacks, where an insider participant may upload poisoning updates to the central entity so that it can cause a construction failure of the global model and thus affect its global performance. Therefore, it is crucial to design security means to detect and mitigate such threats. In this paper, we design a novel framework to automatically detect malicious participants in the FL process. In particular, our framework first uses a deep reinforcement algorithm to dynamically select a network slice as a trusted participant, based mainly on its reputation. The selected participant will then be in charge of identifying poisoning model updates by leveraging unsupervised machine learning. We demonstrate the feasibility of our framework on top of a real dataset that we generate using the 5G OpenAirInterface (OAI) platform. Evaluation results show the efficiency of our framework in dealing with poisoning attacks even with the presence of several malicious participants.