Robust and privacy-preserving federated learning scheme based on ciphertext-selected users

Xiaoming Wang, Zhiquan Liu, Binrui Huang

Abstract: Highlights•A ciphertext-based user selection algorithm is proposed within the framework of federated Learning, which considers both the size and distribution of users' datasets. The algorithm aims to select users with large and uniformly distributed datasets to participate in the training process of the global model, while ensuring that users' data privacy is not compromised. Additionally, it filters out anomalous model parameters with significant differences from the global model parameters without revealing the privacy of users' local model parameters, thus enhancing the robustness of the global model.•To reduce network latency and protect user data privacy, a lightweight encryption algorithm is designed based on cloud-edge collaboration. This algorithm not only effectively protects the privacy of users' local model parameters but also ensures their security during the aggregation process. Moreover, it has high computational and communication efficiency.•Based on the difficulty of the Decisional Diffie-Hellman (DDH)) problem, the proposed scheme is proven to be secure and capable of effectively resisting inference attacks, Byzantine attacks, and collusion attacks, ensuring the global model's robustness. Furthermore, the performance analysis of the proposed scheme, along with a series of related experiments conducted on the MNIST and CIFAR-10 datasets, demonstrates that it outperforms existing schemes in terms of model accuracy and efficiency.