Vertical Federated Learning Backdoor Results Collapse Under Realistic Constraints; BVBench Released to Reset the Field

arXiv 2608.12962·low signalXBlueskyLinkedInCopy link

Prior work reports near-perfect backdoor attack success rates against Vertical Federated Learning and proposes effective defenses, but this systematic study finds most of those findings fail to hold under realistic conditions because existing approaches assume unrealistic prior knowledge and the gap was concealed by poorly designed evaluation practices. The authors redefine threat models under practical constraints, propose realistic backdoor workflows, and release BVBench, a backdoor-centric benchmark preloaded with state-of-the-art baselines. The takeaway for anyone deploying cross-organization VFL: current published risk estimates — in both directions, attack and defense — are not load-bearing.