Abstract

With the rapid development of artificial intelligence, demand for computing power and training data has reached unprecedented levels. Vertical Federated Learning, a promising distributed training method, can effectively address data shortage problems within a single organization. In addition, the emergence of computing power networks (CPNs) provides an ideal platform to address both runtime resource limitations and distributed training issues, also in complex scenarios characterizing modern cyber–physical systems. However, native VFL is susceptible to label, gradient, and model inference attacks. Therefore, to address all of the above challenges, we propose PVFL-CPN, a privacy-preserving VFL method for computing power networks based on homomorphic encryption and secret sharing. This approach uses masking-based one-time pad encryption to secure the forward and backward propagation processes in VFL, as well as to protect communication between computing nodes in CPN. PVFL-CPN outperformed its VFL competitors in terms of accuracy, computational cost, and communication efficiency.