APDPFL: Anti-Poisoning Attack Decentralized Privacy Enhanced Federated Learning Scheme for Flight Operation Data Sharing

Abstract

The sharing of flight operation data brings huge benefits to all participants, but for the privacy protection and data security, it is difficult to directly share flight operation data. Federated learning (FL) enables participants to jointly train machine learning models without exposing local data. However, due to the centralization of FL and the unreliability of FL participants, FL is vulnerable to malicious client and server attacks. In this paper, an anti-poisoning attack decentralized privacy enhanced federated learning (APDPFL) scheme is designed to mitigate the impact of server and malicious clients. Specifically, a local Rényi differential privacy is designed to protect client data privacy. Then, a verification method based on K-Means clustering is proposed to select models to participate in aggregation, which improves the anti-poisoning attack performance. Finally, a federated grouping practical Byzantine fault tolerance (FGPBFT) consensus algorithm based on consortium blockchain is proposed to dynamically change server and consensus clients, to decentralize server and improve the consensus efficiency. The theoretical analysis proves that the APDPFL achieves better convergence and provides data privacy protection and security protection. The experimental results on public datasets and flight operation datasets show that the APDPFL is robust and effective for sharing flight operation data.

Topics & Concepts