Efficient Privacy-Preserving Federated Learning With Integrity in Fog Computing

Federated learning allows edge devices to provide local updates to the cloud server to collaboratively train a global model. Fog computing is a hierarchical computing architecture that adds fog nodes between edge devices and the cloud server to reduce their burden. Recent works apply fog computing to federated learning, improving overall efficiency. However, in fog computing-based federated learning, unprotected updates leak privacy, while low-quality updates compromise model integrity. Therefore, several works have integrated low-quality update processing methods with secure computation techniques to address both problems, but they still encounter challenges in two aspects: (i) efficiency in practice; and (ii) effectiveness of processing low-quality updates, especially in non-IID dataset settings. Aiming at these challenges, this work proposes a privacy-preserving fog computing-based federated learning scheme, named EPIFL, which Efficiently preserves Privacy and Integrity of Federated Learning in fog computing. Specifically, we design an effective low-quality update processing method based on selecting and clipping, named SaC. Then, considering the privacy preservation of updates and the SaC method, we customize secure multi-party computation protocols based on replicated secret sharing, which are robust against fog node dropouts. Experimental results show that EPIFL outperforms existing schemes in mitigating the impact of low-quality updates, and has better computational and communication overheads.