Privacy-preserving vertical federated inference for time series

The paper proposes PPHH-VFL, which accelerates privacy-preserving vertical federated inference for time series by up to six orders of magnitude compared with end-to-end MPC. The architecture splits the model head into a plaintext public head and a lightweight MPC private head, and the public embeddings are adversarially trained to mitigate privacy leakage. On models with up to 86M parameters, it achieves up to 44.4x speedup under WAN and 91.2x lower communication cost than the VFL+MPC baseline, with classification accuracy improved by 2.50% and regression RMSE improved by 40.7%.