Verifiable and Privacy-Preserving Cooperative Federated Learning in UAV-Assisted Vehicular Networks

Federated learning (FL) is a promising distributed learning paradigm, which enables devices to collaboratively train an AI model without exposing participants' private data. However, FL is vulnerable to various attacks and thus remains exposed to privacy issues. For example, malicious parties can launch attacks to recover sensitive and private training data from the shared parameters. Leakage of privacy data can cause serious damage to data providers. Furthermore, user anonymity and data verification in FL also need to be considered. To tackle these problems, in this paper, a verifiable and privacy-preserving cooperative FL (VPPFL) scheme is proposed in UAV-assisted vehicular networks (UVNs). Specifically, to preserve the identity privacy of vehicles, elliptic curve cryptosystem (ECC) is used to generate pseudonyms for vehicles. To preserve the data privacy, Paillier homomorphic encryption algorithm is utilized to encrypt the updates of vehicles, whereby UAVs directly perform global aggregations on encrypted updates instead of raw ones. Additionally, pseudonym-based signature mechanism is presented for vehicles to generate verifiable signatures, so as to ensure the authenticity and validity of uploaded local model updates. Besides, to sufficiently use the multi-source data, multiple UAVs share the local updates packets with each other to execute global aggregation. Finally, simulations are carried out to demonstrate that the proposed scheme can achieve high accuracy and verification with providing strict privacy protection.