Differential Privacy in Vertical Federated Learning With Perturbed Gradients Obtained via Functional Mechanism
Year: 2026
Type: article
Abstract: Vertical federated learning enables multiple institutions to collaboratively train models without sharing raw features, thereby addressing data-silo challenges while satisfying privacy-compliance requirements. Existing differentially private vertical federated learning methods typically inject noise directly into exchanged gradients and embeddings, but under strong privacy constraints they suffer from degraded model accuracy, unstable convergence, and excessive communication overhead. To address these issues, this paper proposes FMVFed, a differentially private vertical federated learning framework based on the functional mechanism, which effectively balances privacy protection, model utility, and communication efficiency. FMVFed achieves differential privacy by injecting Gaussian noise into the coefficient space of the training objective, thereby avoiding the utility loss caused by directly perturbing high-dimensional vectors. Meanwhile, it introduces an adaptive random Top-$k$Sparsification communication strategy to significantly reduce cross-party communication overhead. Experimental results show that even under a small privacy budget (e.g.,$\varepsilon =1$), FMVFed outperforms methods that directly perturb algorithm outputs by 13.47%, 3.13%, and 2.43% on CIFAR-10, FashionMNIST, and MNIST, respectively.
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
Estimated APC paid: $2,645
OpenAlex ID: W7204192043
