Accelerate Vertical Federated Adversarial Learning with Dual-level Decoupled Backpropagation
Keywords: Vertical Federated Learning, Adversarial Training, Delayed Gradient
Abstract: Vertical Federated Learning (VFL) involves multiple participants collaborating to train models on distinct feature sets from the same data samples.
The distributed deployment of VFL models renders them vulnerable to adversarial perturbations during inference, motivating the need to visit the VFL robustness problem.
Adversarial Training (AT) is the predominant approach for enhancing model robustness.
However, its application in VFL, termed Vertical Federated Adversarial Learning (VFAL), faces significant computational challenges:
Generating adversarial examples in AT requires *iterative full propagations across participants with heavy computation overload*, resulting in VFAL training time far exceeding those of regular VFLs.
To address this challenge, we propose ***DecVFAL***, an accelerated **VFAL** framework through a novel **Dec**oupled backpropagation incorporating a *dual-level decoupled mechanism to enable lazy sequential and decoupled parallel backpropagation*.
Lazy sequential backpropagation sequentially updates the adversarial example using timely partial derivatives with respect to the bottom module and delayed partial derivatives for the remaining modules.
Decoupled parallel backpropagation updates these delayed partial derivatives by utilizing module-wise delayed gradients, enabling asynchronous parallel backpropagation with flexible partitions that align with VFL's distributed deployment.
Rigorous theoretical analysis demonstrates that despite introducing multi-source approximate gradients due to the dual decoupled mechanism and the techniques from the existing VFL methods, *DecVFAL* achieves a $\mathcal{O}(1 / \sqrt{\mathcal{K}})$ convergence rate after $\mathcal{K}$ iterations, on par with regular VFL systems.
Experimental results show that, compared to existing methods, *DecVFAL* ensures competitive robustness while significantly achieving about $3\sim10$ times speed up on various datasets.
Primary Area: alignment, fairness, safety, privacy, and societal considerations
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 7434
Loading