Backdoor Vulnerabilities in VFL: Bridging Research and Practice.

Vertical Federated Learning (VFL) allows multiple organizations to collaboratively train AI models using complementary data features without exposing their raw datasets, which is crucial for privacy. However, this asymmetric information structure also creates vulnerabilities, particularly to backdoor attacks. Malicious participants can inject poisoned data during training and activate these backdoors during inference to manipulate predictions. Despite numerous academic studies reporting high attack success rates and effective defenses, this research argues that most findings do not hold under realistic conditions, highlighting a fundamental disconnect between research and practical application. The paper systematically investigates VFL backdoor vulnerabilities, pointing out flaws in methodological design and evaluation practices in existing literature. It shows that current approaches often overlook practical constraints and rely on unrealistic prior knowledge. To bridge this gap, the authors redefine threat models, propose practical backdoor workflows, and introduce BVBench, a new benchmark designed for fair, practical, and comprehensive evaluation of VFL backdoor risks and defenses, revealing the fragility of current understanding.