Robust attack efficacy across diverse VFL environments

Establish robust efficacy for vertical federated learning backdoor attacks across diverse realistic VFL environments and datasets, rather than achieving strong performance only under particular settings.

Background

The paper evaluates existing VFL backdoor attacks under BVBench, which standardizes datasets, architectures, configurations, and evaluation metrics. The results show that attacks often perform well on CIFAR-10 but degrade substantially on realistic VFL datasets, with large variation in attack success rates across datasets.

Because no evaluated attack maintains consistently strong performance across the benchmarked environments, the paper identifies robustness of attack efficacy as an unresolved research problem. Solving it would require attacks that remain effective despite realistic differences in data distributions, model configurations, and VFL deployment conditions.

References

Robust attack efficacy across diverse VFL environments remains largely an open problem.

— Understanding Backdoor Vulnerabilities in Vertical Federated Learning: The Gap Between Research and Practice  (2608.12962 - Zhao et al., 13 Aug 2026) in Section 5, subsection “Efficacy” (Empirical Analysis: Attacks)

Understanding how attacks behave under realistic mixtures of benign and adversarial peers remains an important open question.

— Understanding Backdoor Vulnerabilities in Vertical Federated Learning: The Gap Between Research and Practice  (2608.12962 - Zhao et al., 13 Aug 2026) in Section 5, subsection “Robustness to Adversarial Environment” (Empirical Analysis: Attacks)

Whether future methods can achieve a substantially better efficacy-stealthiness trade-off remains an open question.

— Understanding Backdoor Vulnerabilities in Vertical Federated Learning: The Gap Between Research and Practice  (2608.12962 - Zhao et al., 13 Aug 2026) in Section 5, subsection “Stealthiness” (Empirical Analysis: Attacks)