VFL-Searcher: Optimizing Stealthy Adversarial Dominating Inputs

Vertical Federated Learning (VFL) preserves privacy by splitting features across participants, yet the exchange of intermediate embeddings creates a unique attack surface. We investigate VFL security through Search-Based Software Engineering (SBSE) and propose VFL-Searcher, a search-based framework for systematically auditing the robustness of VFL. Our methodology treats adversarial synthesis as a multi-objective search problem, allowing researchers to automatically discover the Pareto-optimal trade-offs between attack impact and bypass capability. Unlike gradient-based methods, our approach treats the attack as a multi-objective optimization problem, simultaneously maximizing classification impact and minimizing detectability against anomaly detectors. Evaluation across four datasets and five detection algorithms shows that VFL-Searcher achieves near-perfect success in undefended settings and maintains significantly higher stealth against robust detectors compared to baseline attacks.