Trustworthiness in Vertical Federated Learning: Privacy attacks, Robustness Vulnerabilities, and Secure Protocols

Vertical Federated Learning (VFL) enables organizations with distinct features of the same data samples to collaboratively train machine learning models without directly sharing raw data. While VFL offers strong potential for privacy-preserving analytics in domains such as finance and healthcare, its trustworthiness remains insufficiently understood. This thesis provides a systematic study of VFL trustworthiness along three dimensions: privacy, robustness, and secure protocol design. First, we investigate privacy vulnerabilities by introducing URVFL, an undetectable data reconstruction attack. By generating malicious gradients through a discriminator with auxiliary classifier, URVFL achieves state-of-the-art reconstruction fidelity while evading existing detection mechanisms, revealing the fragility of current defenses. Second, we analyze robustness vulnerabilities by examining adversarial example attacks in VFL inference. We formulate adaptive corruption of client-server channels as an online optimization problem and propose the E-TS algorithm, which leverages empirical Thompson sampling to identify optimal corruption patterns with provable regret bounds. Our results demonstrate that adaptive adversaries can severely compromise VFL robustness. Finally, we design FedVS, a synchronous split VFL protocol that simultaneously provides straggler resilience and information-theoretic privacy. Through coding-based secret sharing, FedVS losslessly aggregates embeddings even with stragglers while preventing leakage to curious servers and colluding clients. This thesis demonstrates that trustworthy VFL requires a holistic approach addressing privacy, robustness, and protocol-level vulnerabilities simultaneously. The proposed attacks expose critical weaknesses that must be addressed before large-scale deployment, while our secure protocol provides a constructive foundation for privacy-preserving collaborative learning. These contributions advance understanding of trustworthiness in distributed machine learning and establish principled design criteria for secure VFL systems.