A Federated Learning Scheme with Adaptive Hierarchical Protection and Multiple Aggregation

The results of federated learning contain certain private information and still have the risk of privacy leakage. Therefore, this paper proposes a Federated Learning Scheme with Adaptive Hierarchical Protection and Multiple Aggregation (ADDFed). The scheme trains models and generates updates using local datasets at the client side. Based on comparisons between client contributions and two thresholds, it chooses homomorphic encryption, adaptive differential privacy, or data discarding methods to protect the data. The first aggregation model and the second aggregation model are obtained by classification aggregation at the server side. Finally, the client decrypts and aggregates to get the global model. Experimental results demonstrate that ADDFed consumes a privacy budget of 3624.47, involves 562,219,358 bytes of communication, and requires 11,603 seconds of training time, achieving a model accuracy of 87.4%, thus showing significant advantages.