SABH-FL: a secure attribute-based hierarchical federated learning

Federated learning has recently attracted much attention as it enables multiple clients to cooperatively train a model without uploading sensitive data. Secure aggregation of distributed training local models preserves the privacy of gradients and hence protects the local sensitive data. This paper proposes a secure attribute-based hierarchical federated learning (SABH-FL) approach where the global model is divided into multiple sub-models with hierarchical access and distributed training supported. We propose to protect the local training updates using an Attribute-Based Paillier Encryption technique, ensuring that the different sub-models can be accessed by clients with different attributes and privileges and the encrypted local training results can be aggregated by the server. We also evaluated performance in training accuracy on the MNIST and CIFAR-10 datasets. The results showed that the increase in the number of clients participating with the global model leads to higher training accuracy.