Type-2 Fuzzy Controller as a Client Reputation Mechanism for Federated Learning

Security in federated learning is a critical issue due to the system’s distributed nature. The centralized model aggregates trained models from multiple clients, each using private data. This paper proposes extending this approach by introducing an additional mechanism to assess the client’s reputation based on the model sent. This mechanism utilizes a type-2 fuzzy controller that employs Gaussian functions to account for uncertainties such as potential attacks and poor adaptation to data. The controller’s design is based on the model’s features, including historical similarity and current evaluation metrics. Selected information is fuzzified and processed through fuzzy inference to assess the quality of the given model, which reflects the current reputation level of the client submitting it. The fuzzy logic-based mechanism enables effective reputation assessment, considering uncertainty, which ultimately enhances the security of the federated learning system.